Spaces:
Running
Running
add standalone modeling file
Browse files- dalle_mini/configuration_bart.py +185 -0
- dalle_mini/modeling_bart_flax.py +1023 -0
dalle_mini/configuration_bart.py
ADDED
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2021 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" BART model configuration """
|
16 |
+
import warnings
|
17 |
+
|
18 |
+
from transformers.configuration_utils import PretrainedConfig
|
19 |
+
from transformers.utils import logging
|
20 |
+
|
21 |
+
|
22 |
+
logger = logging.get_logger(__name__)
|
23 |
+
|
24 |
+
BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
25 |
+
"facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/config.json",
|
26 |
+
# See all BART models at https://huggingface.co/models?filter=bart
|
27 |
+
}
|
28 |
+
|
29 |
+
|
30 |
+
class BartConfig(PretrainedConfig):
|
31 |
+
r"""
|
32 |
+
This is the configuration class to store the configuration of a :class:`~transformers.BartModel`. It is used to
|
33 |
+
instantiate a BART model according to the specified arguments, defining the model architecture. Instantiating a
|
34 |
+
configuration with the defaults will yield a similar configuration to that of the BART `facebook/bart-large
|
35 |
+
<https://huggingface.co/facebook/bart-large>`__ architecture.
|
36 |
+
|
37 |
+
Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
|
38 |
+
outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
|
39 |
+
|
40 |
+
|
41 |
+
Args:
|
42 |
+
vocab_size (:obj:`int`, `optional`, defaults to 50265):
|
43 |
+
Vocabulary size of the BART model. Defines the number of different tokens that can be represented by the
|
44 |
+
:obj:`inputs_ids` passed when calling :class:`~transformers.BartModel` or
|
45 |
+
:class:`~transformers.TFBartModel`.
|
46 |
+
d_model (:obj:`int`, `optional`, defaults to 1024):
|
47 |
+
Dimensionality of the layers and the pooler layer.
|
48 |
+
encoder_layers (:obj:`int`, `optional`, defaults to 12):
|
49 |
+
Number of encoder layers.
|
50 |
+
decoder_layers (:obj:`int`, `optional`, defaults to 12):
|
51 |
+
Number of decoder layers.
|
52 |
+
encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
|
53 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
54 |
+
decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
|
55 |
+
Number of attention heads for each attention layer in the Transformer decoder.
|
56 |
+
decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
|
57 |
+
Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
|
58 |
+
encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
|
59 |
+
Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
|
60 |
+
activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
|
61 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string,
|
62 |
+
:obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
|
63 |
+
dropout (:obj:`float`, `optional`, defaults to 0.1):
|
64 |
+
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
65 |
+
attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
66 |
+
The dropout ratio for the attention probabilities.
|
67 |
+
activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
68 |
+
The dropout ratio for activations inside the fully connected layer.
|
69 |
+
classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
|
70 |
+
The dropout ratio for classifier.
|
71 |
+
max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
|
72 |
+
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
73 |
+
just in case (e.g., 512 or 1024 or 2048).
|
74 |
+
init_std (:obj:`float`, `optional`, defaults to 0.02):
|
75 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
76 |
+
encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
|
77 |
+
The LayerDrop probability for the encoder. See the `LayerDrop paper <see
|
78 |
+
https://arxiv.org/abs/1909.11556>`__ for more details.
|
79 |
+
decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
|
80 |
+
The LayerDrop probability for the decoder. See the `LayerDrop paper <see
|
81 |
+
https://arxiv.org/abs/1909.11556>`__ for more details.
|
82 |
+
gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
83 |
+
If True, use gradient checkpointing to save memory at the expense of slower backward pass.
|
84 |
+
scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
|
85 |
+
Scale embeddings by diving by sqrt(d_model).
|
86 |
+
use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
|
87 |
+
Whether or not the model should return the last key/values attentions (not used by all models).
|
88 |
+
num_labels: (:obj:`int`, `optional`, defaults to 3):
|
89 |
+
The number of labels to use in :class:`~transformers.BartForSequenceClassification`.
|
90 |
+
forced_eos_token_id (:obj:`int`, `optional`, defaults to 2):
|
91 |
+
The id of the token to force as the last generated token when :obj:`max_length` is reached. Usually set to
|
92 |
+
:obj:`eos_token_id`.
|
93 |
+
|
94 |
+
Example::
|
95 |
+
|
96 |
+
>>> from transformers import BartModel, BartConfig
|
97 |
+
|
98 |
+
>>> # Initializing a BART facebook/bart-large style configuration
|
99 |
+
>>> configuration = BartConfig()
|
100 |
+
|
101 |
+
>>> # Initializing a model from the facebook/bart-large style configuration
|
102 |
+
>>> model = BartModel(configuration)
|
103 |
+
|
104 |
+
>>> # Accessing the model configuration
|
105 |
+
>>> configuration = model.config
|
106 |
+
"""
|
107 |
+
model_type = "bart"
|
108 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
109 |
+
attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
|
110 |
+
|
111 |
+
def __init__(
|
112 |
+
self,
|
113 |
+
vocab_size=50265,
|
114 |
+
decoder_vocab_size=16384 + 1, # encoded image token space + 1 for bos
|
115 |
+
max_position_embeddings=1024,
|
116 |
+
decoder_max_position_embeddings=256 + 1, # number of encoded tokens + 1 for bos,
|
117 |
+
encoder_layers=12,
|
118 |
+
encoder_ffn_dim=4096,
|
119 |
+
encoder_attention_heads=16,
|
120 |
+
decoder_layers=12,
|
121 |
+
decoder_ffn_dim=4096,
|
122 |
+
decoder_attention_heads=16,
|
123 |
+
encoder_layerdrop=0.0,
|
124 |
+
decoder_layerdrop=0.0,
|
125 |
+
activation_function="gelu",
|
126 |
+
d_model=1024,
|
127 |
+
dropout=0.1,
|
128 |
+
attention_dropout=0.0,
|
129 |
+
activation_dropout=0.0,
|
130 |
+
init_std=0.02,
|
131 |
+
classifier_dropout=0.0,
|
132 |
+
scale_embedding=False,
|
133 |
+
gradient_checkpointing=False,
|
134 |
+
use_cache=True,
|
135 |
+
num_labels=3,
|
136 |
+
pad_token_id=1,
|
137 |
+
bos_token_id=0,
|
138 |
+
eos_token_id=2,
|
139 |
+
is_encoder_decoder=True,
|
140 |
+
decoder_start_token_id=16384,
|
141 |
+
forced_eos_token_id=2,
|
142 |
+
**kwargs,
|
143 |
+
):
|
144 |
+
self.vocab_size = vocab_size
|
145 |
+
self.decoder_vocab_size = decoder_vocab_size
|
146 |
+
self.max_position_embeddings = max_position_embeddings
|
147 |
+
self.decoder_max_position_embeddings = decoder_max_position_embeddings
|
148 |
+
self.d_model = d_model
|
149 |
+
self.encoder_ffn_dim = encoder_ffn_dim
|
150 |
+
self.encoder_layers = encoder_layers
|
151 |
+
self.encoder_attention_heads = encoder_attention_heads
|
152 |
+
self.decoder_ffn_dim = decoder_ffn_dim
|
153 |
+
self.decoder_layers = decoder_layers
|
154 |
+
self.decoder_attention_heads = decoder_attention_heads
|
155 |
+
self.dropout = dropout
|
156 |
+
self.attention_dropout = attention_dropout
|
157 |
+
self.activation_dropout = activation_dropout
|
158 |
+
self.activation_function = activation_function
|
159 |
+
self.init_std = init_std
|
160 |
+
self.encoder_layerdrop = encoder_layerdrop
|
161 |
+
self.decoder_layerdrop = decoder_layerdrop
|
162 |
+
self.classifier_dropout = classifier_dropout
|
163 |
+
self.use_cache = use_cache
|
164 |
+
self.num_hidden_layers = encoder_layers
|
165 |
+
self.gradient_checkpointing = gradient_checkpointing
|
166 |
+
self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
|
167 |
+
|
168 |
+
super().__init__(
|
169 |
+
num_labels=num_labels,
|
170 |
+
pad_token_id=pad_token_id,
|
171 |
+
bos_token_id=bos_token_id,
|
172 |
+
eos_token_id=eos_token_id,
|
173 |
+
is_encoder_decoder=is_encoder_decoder,
|
174 |
+
decoder_start_token_id=decoder_start_token_id,
|
175 |
+
forced_eos_token_id=forced_eos_token_id,
|
176 |
+
**kwargs,
|
177 |
+
)
|
178 |
+
|
179 |
+
# ensure backward compatibility for BART CNN models
|
180 |
+
if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
|
181 |
+
self.forced_bos_token_id = self.bos_token_id
|
182 |
+
warnings.warn(
|
183 |
+
f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions."
|
184 |
+
"The config can simply be saved and uploaded again to be fixed."
|
185 |
+
)
|
dalle_mini/modeling_bart_flax.py
ADDED
@@ -0,0 +1,1023 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2021 The Fairseq Authors and The Google Flax Team Authors And The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" Flax Bart model. """
|
16 |
+
|
17 |
+
import math
|
18 |
+
from functools import partial
|
19 |
+
from typing import Callable, Optional, Tuple
|
20 |
+
|
21 |
+
import numpy as np
|
22 |
+
|
23 |
+
import flax.linen as nn
|
24 |
+
import jax
|
25 |
+
import jax.numpy as jnp
|
26 |
+
from flax.core.frozen_dict import FrozenDict, unfreeze
|
27 |
+
from flax.linen import combine_masks, make_causal_mask
|
28 |
+
from flax.linen.attention import dot_product_attention_weights
|
29 |
+
from jax import lax
|
30 |
+
from jax.random import PRNGKey
|
31 |
+
|
32 |
+
|
33 |
+
from transformers.modeling_flax_outputs import (
|
34 |
+
FlaxBaseModelOutput,
|
35 |
+
FlaxBaseModelOutputWithPastAndCrossAttentions,
|
36 |
+
FlaxCausalLMOutputWithCrossAttentions,
|
37 |
+
FlaxSeq2SeqLMOutput,
|
38 |
+
FlaxSeq2SeqModelOutput,
|
39 |
+
)
|
40 |
+
from transformers.modeling_flax_utils import (
|
41 |
+
ACT2FN,
|
42 |
+
FlaxPreTrainedModel,
|
43 |
+
)
|
44 |
+
from transformers.utils import logging
|
45 |
+
|
46 |
+
|
47 |
+
from configuration_bart import BartConfig
|
48 |
+
|
49 |
+
|
50 |
+
logger = logging.get_logger(__name__)
|
51 |
+
|
52 |
+
|
53 |
+
def shift_tokens_right(input_ids: np.array, pad_token_id: int, decoder_start_token_id: int) -> np.ndarray:
|
54 |
+
"""
|
55 |
+
Shift input ids one token to the right.
|
56 |
+
"""
|
57 |
+
shifted_input_ids = np.zeros_like(input_ids)
|
58 |
+
shifted_input_ids[:, 1:] = input_ids[:, :-1]
|
59 |
+
shifted_input_ids[:, 0] = decoder_start_token_id
|
60 |
+
|
61 |
+
shifted_input_ids = np.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
|
62 |
+
return shifted_input_ids
|
63 |
+
|
64 |
+
|
65 |
+
class FlaxBartAttention(nn.Module):
|
66 |
+
config: BartConfig
|
67 |
+
embed_dim: int
|
68 |
+
num_heads: int
|
69 |
+
dropout: float = 0.0
|
70 |
+
causal: bool = False
|
71 |
+
bias: bool = True
|
72 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
73 |
+
|
74 |
+
def setup(self) -> None:
|
75 |
+
self.head_dim = self.embed_dim // self.num_heads
|
76 |
+
assert (
|
77 |
+
self.head_dim * self.num_heads == self.embed_dim
|
78 |
+
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
|
79 |
+
|
80 |
+
dense = partial(
|
81 |
+
nn.Dense,
|
82 |
+
self.embed_dim,
|
83 |
+
use_bias=self.bias,
|
84 |
+
dtype=self.dtype,
|
85 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
86 |
+
)
|
87 |
+
|
88 |
+
self.q_proj, self.k_proj, self.v_proj = dense(), dense(), dense()
|
89 |
+
self.out_proj = dense()
|
90 |
+
|
91 |
+
self.dropout_layer = nn.Dropout(rate=self.dropout)
|
92 |
+
|
93 |
+
if self.causal:
|
94 |
+
self.causal_mask = make_causal_mask(
|
95 |
+
jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
|
96 |
+
)
|
97 |
+
|
98 |
+
def _split_heads(self, hidden_states):
|
99 |
+
return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
|
100 |
+
|
101 |
+
def _merge_heads(self, hidden_states):
|
102 |
+
return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
|
103 |
+
|
104 |
+
@nn.compact
|
105 |
+
def _concatenate_to_cache(self, key, value, query, attention_mask):
|
106 |
+
"""
|
107 |
+
This function takes projected key, value states from a single input token and concatenates the states to cached
|
108 |
+
states from previous steps. This function is slighly adapted from the official Flax repository:
|
109 |
+
https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
|
110 |
+
"""
|
111 |
+
# detect if we're initializing by absence of existing cache data.
|
112 |
+
is_initialized = self.has_variable("cache", "cached_key")
|
113 |
+
cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
|
114 |
+
cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
|
115 |
+
cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
|
116 |
+
|
117 |
+
if is_initialized:
|
118 |
+
*batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
|
119 |
+
# update key, value caches with our new 1d spatial slices
|
120 |
+
cur_index = cache_index.value
|
121 |
+
indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
|
122 |
+
key = lax.dynamic_update_slice(cached_key.value, key, indices)
|
123 |
+
value = lax.dynamic_update_slice(cached_value.value, value, indices)
|
124 |
+
cached_key.value = key
|
125 |
+
cached_value.value = value
|
126 |
+
num_updated_cache_vectors = query.shape[1]
|
127 |
+
cache_index.value = cache_index.value + num_updated_cache_vectors
|
128 |
+
# causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
|
129 |
+
pad_mask = jnp.broadcast_to(
|
130 |
+
jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
|
131 |
+
tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
|
132 |
+
)
|
133 |
+
attention_mask = combine_masks(pad_mask, attention_mask)
|
134 |
+
return key, value, attention_mask
|
135 |
+
|
136 |
+
def __call__(
|
137 |
+
self,
|
138 |
+
hidden_states: jnp.ndarray,
|
139 |
+
attention_mask: jnp.ndarray,
|
140 |
+
key_value_states: Optional[jnp.ndarray] = None,
|
141 |
+
init_cache: bool = False,
|
142 |
+
deterministic: bool = True,
|
143 |
+
) -> Tuple[jnp.ndarray]:
|
144 |
+
"""Input shape: Batch x Time x Channel"""
|
145 |
+
|
146 |
+
# if key_value_states are provided this layer is used as a cross-attention layer
|
147 |
+
# for the decoder
|
148 |
+
is_cross_attention = key_value_states is not None
|
149 |
+
batch_size = hidden_states.shape[0]
|
150 |
+
|
151 |
+
# get query proj
|
152 |
+
query_states = self.q_proj(hidden_states)
|
153 |
+
# get key, value proj
|
154 |
+
if is_cross_attention:
|
155 |
+
# cross_attentions
|
156 |
+
key_states = self.k_proj(key_value_states)
|
157 |
+
value_states = self.v_proj(key_value_states)
|
158 |
+
else:
|
159 |
+
# self_attention
|
160 |
+
key_states = self.k_proj(hidden_states)
|
161 |
+
value_states = self.v_proj(hidden_states)
|
162 |
+
|
163 |
+
query_states = self._split_heads(query_states)
|
164 |
+
key_states = self._split_heads(key_states)
|
165 |
+
value_states = self._split_heads(value_states)
|
166 |
+
|
167 |
+
# handle cache prepare causal attention mask
|
168 |
+
if self.causal:
|
169 |
+
query_length, key_length = query_states.shape[1], key_states.shape[1]
|
170 |
+
if self.has_variable("cache", "cached_key"):
|
171 |
+
mask_shift = self.variables["cache"]["cache_index"]
|
172 |
+
max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
|
173 |
+
causal_mask = lax.dynamic_slice(
|
174 |
+
self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
|
175 |
+
)
|
176 |
+
else:
|
177 |
+
causal_mask = self.causal_mask[:, :, :query_length, :key_length]
|
178 |
+
causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
|
179 |
+
|
180 |
+
# combine masks if needed
|
181 |
+
if self.causal:
|
182 |
+
attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
|
183 |
+
attention_mask = combine_masks(attention_mask, causal_mask)
|
184 |
+
else:
|
185 |
+
attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
|
186 |
+
|
187 |
+
# During fast autoregressive decoding, we feed one position at a time,
|
188 |
+
# and cache the keys and values step by step.
|
189 |
+
if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
|
190 |
+
key_states, value_states, attention_mask = self._concatenate_to_cache(
|
191 |
+
key_states, value_states, query_states, attention_mask
|
192 |
+
)
|
193 |
+
|
194 |
+
# Convert the boolean attention mask to an attention bias.
|
195 |
+
# attention mask in the form of attention bias
|
196 |
+
attention_bias = lax.select(
|
197 |
+
attention_mask > 0,
|
198 |
+
jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
|
199 |
+
jnp.full(attention_mask.shape, float("-inf")).astype(self.dtype),
|
200 |
+
)
|
201 |
+
|
202 |
+
dropout_rng = None
|
203 |
+
if not deterministic and self.dropout > 0.0:
|
204 |
+
dropout_rng = self.make_rng("dropout")
|
205 |
+
|
206 |
+
attn_weights = dot_product_attention_weights(
|
207 |
+
query_states,
|
208 |
+
key_states,
|
209 |
+
bias=attention_bias,
|
210 |
+
dropout_rng=dropout_rng,
|
211 |
+
dropout_rate=self.dropout,
|
212 |
+
broadcast_dropout=True,
|
213 |
+
deterministic=deterministic,
|
214 |
+
dtype=self.dtype,
|
215 |
+
precision=None,
|
216 |
+
)
|
217 |
+
|
218 |
+
attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
|
219 |
+
attn_output = self._merge_heads(attn_output)
|
220 |
+
attn_output = self.out_proj(attn_output)
|
221 |
+
|
222 |
+
return attn_output
|
223 |
+
|
224 |
+
|
225 |
+
class FlaxBartEncoderLayer(nn.Module):
|
226 |
+
config: BartConfig
|
227 |
+
dtype: jnp.dtype = jnp.float32
|
228 |
+
|
229 |
+
def setup(self) -> None:
|
230 |
+
self.embed_dim = self.config.d_model
|
231 |
+
self.self_attn = FlaxBartAttention(
|
232 |
+
config=self.config,
|
233 |
+
embed_dim=self.embed_dim,
|
234 |
+
num_heads=self.config.encoder_attention_heads,
|
235 |
+
dropout=self.config.attention_dropout,
|
236 |
+
dtype=self.dtype,
|
237 |
+
)
|
238 |
+
self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
239 |
+
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
240 |
+
self.activation_fn = ACT2FN[self.config.activation_function]
|
241 |
+
self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
|
242 |
+
self.fc1 = nn.Dense(
|
243 |
+
self.config.encoder_ffn_dim,
|
244 |
+
dtype=self.dtype,
|
245 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
246 |
+
)
|
247 |
+
self.fc2 = nn.Dense(
|
248 |
+
self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
|
249 |
+
)
|
250 |
+
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
251 |
+
|
252 |
+
def __call__(
|
253 |
+
self,
|
254 |
+
hidden_states: jnp.ndarray,
|
255 |
+
attention_mask: jnp.ndarray,
|
256 |
+
deterministic: bool = True,
|
257 |
+
) -> Tuple[jnp.ndarray]:
|
258 |
+
residual = hidden_states
|
259 |
+
hidden_states = self.self_attn(hidden_states=hidden_states, attention_mask=attention_mask)
|
260 |
+
|
261 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
262 |
+
hidden_states = residual + hidden_states
|
263 |
+
hidden_states = self.self_attn_layer_norm(hidden_states)
|
264 |
+
|
265 |
+
residual = hidden_states
|
266 |
+
hidden_states = self.activation_fn(self.fc1(hidden_states))
|
267 |
+
hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
|
268 |
+
hidden_states = self.fc2(hidden_states)
|
269 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
270 |
+
hidden_states = residual + hidden_states
|
271 |
+
hidden_states = self.final_layer_norm(hidden_states)
|
272 |
+
|
273 |
+
return hidden_states
|
274 |
+
|
275 |
+
|
276 |
+
class FlaxBartEncoderLayerCollection(nn.Module):
|
277 |
+
config: BartConfig
|
278 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
279 |
+
|
280 |
+
def setup(self):
|
281 |
+
self.layers = [
|
282 |
+
FlaxBartEncoderLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.encoder_layers)
|
283 |
+
]
|
284 |
+
|
285 |
+
def __call__(
|
286 |
+
self,
|
287 |
+
hidden_states,
|
288 |
+
attention_mask,
|
289 |
+
deterministic: bool = True,
|
290 |
+
):
|
291 |
+
|
292 |
+
for encoder_layer in self.layers:
|
293 |
+
hidden_states = encoder_layer(
|
294 |
+
hidden_states,
|
295 |
+
attention_mask,
|
296 |
+
deterministic,
|
297 |
+
)
|
298 |
+
|
299 |
+
return FlaxBaseModelOutput(last_hidden_state=hidden_states)
|
300 |
+
|
301 |
+
|
302 |
+
class FlaxBartDecoderLayer(nn.Module):
|
303 |
+
config: BartConfig
|
304 |
+
dtype: jnp.dtype = jnp.float32
|
305 |
+
|
306 |
+
def setup(self) -> None:
|
307 |
+
self.embed_dim = self.config.d_model
|
308 |
+
self.self_attn = FlaxBartAttention(
|
309 |
+
config=self.config,
|
310 |
+
embed_dim=self.embed_dim,
|
311 |
+
num_heads=self.config.decoder_attention_heads,
|
312 |
+
dropout=self.config.attention_dropout,
|
313 |
+
causal=True,
|
314 |
+
dtype=self.dtype,
|
315 |
+
)
|
316 |
+
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
317 |
+
self.activation_fn = ACT2FN[self.config.activation_function]
|
318 |
+
self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
|
319 |
+
|
320 |
+
self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
321 |
+
self.encoder_attn = FlaxBartAttention(
|
322 |
+
config=self.config,
|
323 |
+
embed_dim=self.embed_dim,
|
324 |
+
num_heads=self.config.decoder_attention_heads,
|
325 |
+
dropout=self.config.attention_dropout,
|
326 |
+
dtype=self.dtype,
|
327 |
+
)
|
328 |
+
self.encoder_attn_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
329 |
+
self.fc1 = nn.Dense(
|
330 |
+
self.config.encoder_ffn_dim,
|
331 |
+
dtype=self.dtype,
|
332 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
333 |
+
)
|
334 |
+
self.fc2 = nn.Dense(
|
335 |
+
self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
|
336 |
+
)
|
337 |
+
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
338 |
+
|
339 |
+
def __call__(
|
340 |
+
self,
|
341 |
+
hidden_states: jnp.ndarray,
|
342 |
+
attention_mask: jnp.ndarray,
|
343 |
+
encoder_hidden_states: jnp.ndarray,
|
344 |
+
encoder_attention_mask: Optional[jnp.ndarray] = None,
|
345 |
+
init_cache: bool = False,
|
346 |
+
deterministic: bool = True,
|
347 |
+
) -> Tuple[jnp.ndarray]:
|
348 |
+
residual = hidden_states
|
349 |
+
|
350 |
+
# Self Attention
|
351 |
+
hidden_states = self.self_attn(
|
352 |
+
hidden_states=hidden_states, attention_mask=attention_mask, init_cache=init_cache
|
353 |
+
)
|
354 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
355 |
+
hidden_states = residual + hidden_states
|
356 |
+
hidden_states = self.self_attn_layer_norm(hidden_states)
|
357 |
+
|
358 |
+
# Cross-Attention Block
|
359 |
+
residual = hidden_states
|
360 |
+
|
361 |
+
hidden_states = self.encoder_attn(
|
362 |
+
hidden_states=hidden_states,
|
363 |
+
key_value_states=encoder_hidden_states,
|
364 |
+
attention_mask=encoder_attention_mask,
|
365 |
+
)
|
366 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
367 |
+
hidden_states = residual + hidden_states
|
368 |
+
hidden_states = self.encoder_attn_layer_norm(hidden_states)
|
369 |
+
|
370 |
+
# Fully Connected
|
371 |
+
residual = hidden_states
|
372 |
+
hidden_states = self.activation_fn(self.fc1(hidden_states))
|
373 |
+
hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
|
374 |
+
hidden_states = self.fc2(hidden_states)
|
375 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
376 |
+
hidden_states = residual + hidden_states
|
377 |
+
hidden_states = self.final_layer_norm(hidden_states)
|
378 |
+
|
379 |
+
return hidden_states
|
380 |
+
|
381 |
+
|
382 |
+
class FlaxBartDecoderLayerCollection(nn.Module):
|
383 |
+
config: BartConfig
|
384 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
385 |
+
|
386 |
+
def setup(self):
|
387 |
+
self.layers = [
|
388 |
+
FlaxBartDecoderLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.decoder_layers)
|
389 |
+
]
|
390 |
+
|
391 |
+
def __call__(
|
392 |
+
self,
|
393 |
+
hidden_states,
|
394 |
+
attention_mask,
|
395 |
+
encoder_hidden_states: Optional[jnp.ndarray] = None,
|
396 |
+
encoder_attention_mask: Optional[jnp.ndarray] = None,
|
397 |
+
deterministic: bool = True,
|
398 |
+
init_cache: bool = False,
|
399 |
+
):
|
400 |
+
# decoder layers
|
401 |
+
for decoder_layer in self.layers:
|
402 |
+
hidden_states = decoder_layer(
|
403 |
+
hidden_states,
|
404 |
+
attention_mask=attention_mask,
|
405 |
+
encoder_hidden_states=encoder_hidden_states,
|
406 |
+
encoder_attention_mask=encoder_attention_mask,
|
407 |
+
init_cache=init_cache,
|
408 |
+
deterministic=deterministic,
|
409 |
+
)
|
410 |
+
|
411 |
+
return FlaxBaseModelOutputWithPastAndCrossAttentions(last_hidden_state=hidden_states)
|
412 |
+
|
413 |
+
|
414 |
+
class FlaxBartEncoder(nn.Module):
|
415 |
+
config: BartConfig
|
416 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
417 |
+
embed_tokens: Optional[nn.Embed] = None
|
418 |
+
|
419 |
+
def setup(self):
|
420 |
+
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
421 |
+
|
422 |
+
embed_dim = self.config.d_model
|
423 |
+
self.padding_idx = self.config.pad_token_id
|
424 |
+
self.max_source_positions = self.config.max_position_embeddings
|
425 |
+
self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
|
426 |
+
|
427 |
+
if self.embed_tokens is None:
|
428 |
+
self.embed_tokens = nn.Embed(
|
429 |
+
self.config.vocab_size,
|
430 |
+
embed_dim,
|
431 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
432 |
+
)
|
433 |
+
|
434 |
+
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
435 |
+
# and adjust num_embeddings appropriately. Other models don't have this hack
|
436 |
+
self.offset = 2
|
437 |
+
self.embed_positions = nn.Embed(
|
438 |
+
self.config.max_position_embeddings + self.offset,
|
439 |
+
embed_dim,
|
440 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
441 |
+
)
|
442 |
+
self.layers = FlaxBartEncoderLayerCollection(self.config, self.dtype)
|
443 |
+
self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype)
|
444 |
+
|
445 |
+
def __call__(
|
446 |
+
self,
|
447 |
+
input_ids,
|
448 |
+
attention_mask,
|
449 |
+
position_ids,
|
450 |
+
deterministic: bool = True,
|
451 |
+
):
|
452 |
+
input_shape = input_ids.shape
|
453 |
+
input_ids = input_ids.reshape(-1, input_shape[-1])
|
454 |
+
|
455 |
+
inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
|
456 |
+
|
457 |
+
embed_pos = self.embed_positions(position_ids + self.offset)
|
458 |
+
|
459 |
+
hidden_states = inputs_embeds + embed_pos
|
460 |
+
hidden_states = self.layernorm_embedding(hidden_states)
|
461 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
462 |
+
|
463 |
+
outputs = self.layers(hidden_states, attention_mask, deterministic=deterministic)
|
464 |
+
|
465 |
+
return FlaxBaseModelOutput(
|
466 |
+
last_hidden_state=outputs.last_hidden_state,
|
467 |
+
hidden_states=outputs.hidden_states,
|
468 |
+
attentions=outputs.attentions,
|
469 |
+
)
|
470 |
+
|
471 |
+
|
472 |
+
class FlaxBartDecoder(nn.Module):
|
473 |
+
config: BartConfig
|
474 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
475 |
+
embed_tokens: Optional[nn.Embed] = None
|
476 |
+
|
477 |
+
def setup(self):
|
478 |
+
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
479 |
+
|
480 |
+
embed_dim = self.config.d_model
|
481 |
+
self.padding_idx = self.config.pad_token_id
|
482 |
+
self.max_target_positions = self.config.max_position_embeddings
|
483 |
+
self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
|
484 |
+
|
485 |
+
if self.embed_tokens is None:
|
486 |
+
self.embed_tokens = nn.Embed(
|
487 |
+
self.config.vocab_size,
|
488 |
+
embed_dim,
|
489 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
490 |
+
)
|
491 |
+
|
492 |
+
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
493 |
+
# and adjust num_embeddings appropriately. Other models don't have this hack
|
494 |
+
self.offset = 2
|
495 |
+
self.embed_positions = nn.Embed(
|
496 |
+
self.config.max_position_embeddings + self.offset,
|
497 |
+
embed_dim,
|
498 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
499 |
+
)
|
500 |
+
|
501 |
+
self.layers = FlaxBartDecoderLayerCollection(self.config, self.dtype)
|
502 |
+
self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype)
|
503 |
+
|
504 |
+
def __call__(
|
505 |
+
self,
|
506 |
+
input_ids,
|
507 |
+
attention_mask,
|
508 |
+
position_ids,
|
509 |
+
encoder_hidden_states: Optional[jnp.ndarray] = None,
|
510 |
+
encoder_attention_mask: Optional[jnp.ndarray] = None,
|
511 |
+
init_cache: bool = False,
|
512 |
+
deterministic: bool = True,
|
513 |
+
):
|
514 |
+
input_shape = input_ids.shape
|
515 |
+
input_ids = input_ids.reshape(-1, input_shape[-1])
|
516 |
+
|
517 |
+
inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
|
518 |
+
|
519 |
+
# embed positions
|
520 |
+
positions = self.embed_positions(position_ids + self.offset)
|
521 |
+
|
522 |
+
hidden_states = inputs_embeds + positions
|
523 |
+
hidden_states = self.layernorm_embedding(hidden_states)
|
524 |
+
|
525 |
+
hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
|
526 |
+
|
527 |
+
outputs = self.layers(
|
528 |
+
hidden_states,
|
529 |
+
attention_mask,
|
530 |
+
encoder_hidden_states,
|
531 |
+
encoder_attention_mask,
|
532 |
+
deterministic=deterministic,
|
533 |
+
init_cache=init_cache,
|
534 |
+
)
|
535 |
+
|
536 |
+
return FlaxBaseModelOutputWithPastAndCrossAttentions(
|
537 |
+
last_hidden_state=outputs.last_hidden_state,
|
538 |
+
hidden_states=outputs.hidden_states,
|
539 |
+
attentions=outputs.attentions,
|
540 |
+
cross_attentions=outputs.cross_attentions,
|
541 |
+
)
|
542 |
+
|
543 |
+
|
544 |
+
class FlaxBartModule(nn.Module):
|
545 |
+
config: BartConfig
|
546 |
+
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
547 |
+
|
548 |
+
def setup(self):
|
549 |
+
self.shared = nn.Embed(
|
550 |
+
self.config.vocab_size,
|
551 |
+
self.config.d_model,
|
552 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
553 |
+
)
|
554 |
+
# a separate embedding is used for the decoder
|
555 |
+
self.decoder_embed = nn.Embed(
|
556 |
+
self.config.decoder_vocab_size,
|
557 |
+
self.config.d_model,
|
558 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
559 |
+
)
|
560 |
+
|
561 |
+
self.encoder = FlaxBartEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
|
562 |
+
self.decoder = FlaxBartDecoder(self.config, dtype=self.dtype, embed_tokens=self.decoder_embed)
|
563 |
+
|
564 |
+
def _get_encoder_module(self):
|
565 |
+
return self.encoder
|
566 |
+
|
567 |
+
def _get_decoder_module(self):
|
568 |
+
return self.decoder
|
569 |
+
|
570 |
+
def __call__(
|
571 |
+
self,
|
572 |
+
input_ids,
|
573 |
+
attention_mask,
|
574 |
+
decoder_input_ids,
|
575 |
+
decoder_attention_mask,
|
576 |
+
position_ids,
|
577 |
+
decoder_position_ids,
|
578 |
+
output_attentions: bool = False,
|
579 |
+
output_hidden_states: bool = False,
|
580 |
+
return_dict: bool = True,
|
581 |
+
deterministic: bool = True,
|
582 |
+
):
|
583 |
+
encoder_outputs = self.encoder(
|
584 |
+
input_ids=input_ids,
|
585 |
+
attention_mask=attention_mask,
|
586 |
+
position_ids=position_ids,
|
587 |
+
output_attentions=output_attentions,
|
588 |
+
output_hidden_states=output_hidden_states,
|
589 |
+
return_dict=return_dict,
|
590 |
+
deterministic=deterministic,
|
591 |
+
)
|
592 |
+
|
593 |
+
decoder_outputs = self.decoder(
|
594 |
+
input_ids=decoder_input_ids,
|
595 |
+
attention_mask=decoder_attention_mask,
|
596 |
+
position_ids=decoder_position_ids,
|
597 |
+
encoder_hidden_states=encoder_outputs[0],
|
598 |
+
encoder_attention_mask=attention_mask,
|
599 |
+
output_attentions=output_attentions,
|
600 |
+
output_hidden_states=output_hidden_states,
|
601 |
+
return_dict=return_dict,
|
602 |
+
deterministic=deterministic,
|
603 |
+
)
|
604 |
+
|
605 |
+
if not return_dict:
|
606 |
+
return decoder_outputs + encoder_outputs
|
607 |
+
|
608 |
+
return FlaxSeq2SeqModelOutput(
|
609 |
+
last_hidden_state=decoder_outputs.last_hidden_state,
|
610 |
+
decoder_hidden_states=decoder_outputs.hidden_states,
|
611 |
+
decoder_attentions=decoder_outputs.attentions,
|
612 |
+
cross_attentions=decoder_outputs.cross_attentions,
|
613 |
+
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
|
614 |
+
encoder_hidden_states=encoder_outputs.hidden_states,
|
615 |
+
encoder_attentions=encoder_outputs.attentions,
|
616 |
+
)
|
617 |
+
|
618 |
+
|
619 |
+
class FlaxBartPreTrainedModel(FlaxPreTrainedModel):
|
620 |
+
config_class = BartConfig
|
621 |
+
base_model_prefix: str = "model"
|
622 |
+
module_class: nn.Module = None
|
623 |
+
|
624 |
+
def __init__(
|
625 |
+
self,
|
626 |
+
config: BartConfig,
|
627 |
+
input_shape: Tuple[int] = (1, 1),
|
628 |
+
seed: int = 0,
|
629 |
+
dtype: jnp.dtype = jnp.float32,
|
630 |
+
**kwargs,
|
631 |
+
):
|
632 |
+
module = self.module_class(config=config, dtype=dtype, **kwargs)
|
633 |
+
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
|
634 |
+
|
635 |
+
def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
|
636 |
+
# init input tensors
|
637 |
+
input_ids = jnp.zeros(input_shape, dtype="i4")
|
638 |
+
# make sure initialization pass will work for FlaxBartForSequenceClassificationModule
|
639 |
+
input_ids = jax.ops.index_update(input_ids, (..., -1), self.config.eos_token_id)
|
640 |
+
attention_mask = jnp.ones_like(input_ids)
|
641 |
+
decoder_input_ids = input_ids
|
642 |
+
decoder_attention_mask = jnp.ones_like(input_ids)
|
643 |
+
|
644 |
+
batch_size, sequence_length = input_ids.shape
|
645 |
+
position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
|
646 |
+
decoder_position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
|
647 |
+
|
648 |
+
params_rng, dropout_rng = jax.random.split(rng)
|
649 |
+
rngs = {"params": params_rng, "dropout": dropout_rng}
|
650 |
+
|
651 |
+
return self.module.init(
|
652 |
+
rngs,
|
653 |
+
input_ids,
|
654 |
+
attention_mask,
|
655 |
+
decoder_input_ids,
|
656 |
+
decoder_attention_mask,
|
657 |
+
position_ids,
|
658 |
+
decoder_position_ids,
|
659 |
+
)["params"]
|
660 |
+
|
661 |
+
def init_cache(self, batch_size, max_length, encoder_outputs):
|
662 |
+
r"""
|
663 |
+
Args:
|
664 |
+
batch_size (:obj:`int`):
|
665 |
+
batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
|
666 |
+
max_length (:obj:`int`):
|
667 |
+
maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
|
668 |
+
cache.
|
669 |
+
encoder_outputs (:obj:`Union[FlaxBaseModelOutput, tuple(tuple(jnp.ndarray)]`):
|
670 |
+
``encoder_outputs`` consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`,
|
671 |
+
`optional`: :obj:`attentions`). :obj:`last_hidden_state` of shape :obj:`(batch_size, sequence_length,
|
672 |
+
hidden_size)`, `optional`) is a sequence of hidden-states at the output of the last layer of the
|
673 |
+
encoder. Used in the cross-attention of the decoder.
|
674 |
+
"""
|
675 |
+
# init input variables to retrieve cache
|
676 |
+
decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
|
677 |
+
decoder_attention_mask = jnp.ones_like(decoder_input_ids)
|
678 |
+
decoder_position_ids = jnp.broadcast_to(
|
679 |
+
jnp.arange(jnp.atleast_2d(decoder_input_ids).shape[-1]), decoder_input_ids.shape
|
680 |
+
)
|
681 |
+
|
682 |
+
def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
|
683 |
+
decoder_module = module._get_decoder_module()
|
684 |
+
return decoder_module(
|
685 |
+
decoder_input_ids,
|
686 |
+
decoder_attention_mask,
|
687 |
+
decoder_position_ids,
|
688 |
+
**kwargs,
|
689 |
+
)
|
690 |
+
|
691 |
+
init_variables = self.module.init(
|
692 |
+
jax.random.PRNGKey(0),
|
693 |
+
decoder_input_ids=decoder_input_ids,
|
694 |
+
decoder_attention_mask=decoder_attention_mask,
|
695 |
+
decoder_position_ids=decoder_position_ids,
|
696 |
+
encoder_hidden_states=encoder_outputs[0],
|
697 |
+
init_cache=True,
|
698 |
+
method=_decoder_forward, # we only need to call the decoder to init the cache
|
699 |
+
)
|
700 |
+
return unfreeze(init_variables["cache"])
|
701 |
+
|
702 |
+
def encode(
|
703 |
+
self,
|
704 |
+
input_ids: jnp.ndarray,
|
705 |
+
attention_mask: Optional[jnp.ndarray] = None,
|
706 |
+
position_ids: Optional[jnp.ndarray] = None,
|
707 |
+
train: bool = False,
|
708 |
+
params: dict = None,
|
709 |
+
dropout_rng: PRNGKey = None,
|
710 |
+
):
|
711 |
+
r"""
|
712 |
+
Returns:
|
713 |
+
|
714 |
+
Example::
|
715 |
+
|
716 |
+
>>> from transformers import BartTokenizer, FlaxBartForConditionalGeneration
|
717 |
+
|
718 |
+
>>> model = FlaxBartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')
|
719 |
+
>>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')
|
720 |
+
|
721 |
+
>>> text = "My friends are cool but they eat too many carbs."
|
722 |
+
>>> inputs = tokenizer(text, max_length=1024, return_tensors='jax')
|
723 |
+
>>> encoder_outputs = model.encode(**inputs)
|
724 |
+
"""
|
725 |
+
if attention_mask is None:
|
726 |
+
attention_mask = jnp.ones_like(input_ids)
|
727 |
+
if position_ids is None:
|
728 |
+
batch_size, sequence_length = input_ids.shape
|
729 |
+
position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
|
730 |
+
|
731 |
+
# Handle any PRNG if needed
|
732 |
+
rngs = {}
|
733 |
+
if dropout_rng is not None:
|
734 |
+
rngs["dropout"] = dropout_rng
|
735 |
+
|
736 |
+
def _encoder_forward(module, input_ids, attention_mask, position_ids, **kwargs):
|
737 |
+
encode_module = module._get_encoder_module()
|
738 |
+
return encode_module(input_ids, attention_mask, position_ids, **kwargs)
|
739 |
+
|
740 |
+
return self.module.apply(
|
741 |
+
{"params": params or self.params},
|
742 |
+
input_ids=jnp.array(input_ids, dtype="i4"),
|
743 |
+
attention_mask=jnp.array(attention_mask, dtype="i4"),
|
744 |
+
position_ids=jnp.array(position_ids, dtype="i4"),
|
745 |
+
deterministic=not train,
|
746 |
+
rngs=rngs,
|
747 |
+
method=_encoder_forward,
|
748 |
+
)
|
749 |
+
|
750 |
+
def __call__(
|
751 |
+
self,
|
752 |
+
input_ids: jnp.ndarray,
|
753 |
+
attention_mask: Optional[jnp.ndarray] = None,
|
754 |
+
decoder_input_ids: Optional[jnp.ndarray] = None,
|
755 |
+
decoder_attention_mask: Optional[jnp.ndarray] = None,
|
756 |
+
position_ids: Optional[jnp.ndarray] = None,
|
757 |
+
decoder_position_ids: Optional[jnp.ndarray] = None,
|
758 |
+
output_attentions: Optional[bool] = None,
|
759 |
+
output_hidden_states: Optional[bool] = None,
|
760 |
+
return_dict: Optional[bool] = None,
|
761 |
+
train: bool = False,
|
762 |
+
params: dict = None,
|
763 |
+
dropout_rng: PRNGKey = None,
|
764 |
+
):
|
765 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
766 |
+
output_hidden_states = (
|
767 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
768 |
+
)
|
769 |
+
return_dict = return_dict if return_dict is not None else self.config.return_dict
|
770 |
+
|
771 |
+
# prepare encoder inputs
|
772 |
+
if attention_mask is None:
|
773 |
+
attention_mask = jnp.ones_like(input_ids)
|
774 |
+
if position_ids is None:
|
775 |
+
batch_size, sequence_length = input_ids.shape
|
776 |
+
position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
|
777 |
+
|
778 |
+
# prepare decoder inputs
|
779 |
+
if decoder_input_ids is None:
|
780 |
+
decoder_input_ids = shift_tokens_right(
|
781 |
+
input_ids, self.config.pad_token_id, decoder_start_token_id=self.config.decoder_start_token_id
|
782 |
+
)
|
783 |
+
if decoder_attention_mask is None:
|
784 |
+
decoder_attention_mask = jnp.ones_like(decoder_input_ids)
|
785 |
+
if decoder_position_ids is None:
|
786 |
+
batch_size, sequence_length = decoder_input_ids.shape
|
787 |
+
decoder_position_ids = jnp.broadcast_to(
|
788 |
+
jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
|
789 |
+
)
|
790 |
+
|
791 |
+
# Handle any PRNG if needed
|
792 |
+
rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
|
793 |
+
|
794 |
+
return self.module.apply(
|
795 |
+
{"params": params or self.params},
|
796 |
+
input_ids=jnp.array(input_ids, dtype="i4"),
|
797 |
+
attention_mask=jnp.array(attention_mask, dtype="i4"),
|
798 |
+
position_ids=jnp.array(position_ids, dtype="i4"),
|
799 |
+
decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
|
800 |
+
decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
|
801 |
+
decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
|
802 |
+
deterministic=not train,
|
803 |
+
rngs=rngs,
|
804 |
+
)
|
805 |
+
|
806 |
+
|
807 |
+
class FlaxBartForConditionalGenerationModule(nn.Module):
|
808 |
+
config: BartConfig
|
809 |
+
dtype: jnp.dtype = jnp.float32
|
810 |
+
bias_init: Callable[..., jnp.ndarray] = jax.nn.initializers.zeros
|
811 |
+
|
812 |
+
def setup(self):
|
813 |
+
self.model = FlaxBartModule(config=self.config, dtype=self.dtype)
|
814 |
+
self.lm_head = nn.Dense(
|
815 |
+
self.config.decoder_vocab_size,
|
816 |
+
use_bias=False,
|
817 |
+
dtype=self.dtype,
|
818 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
819 |
+
)
|
820 |
+
self.final_logits_bias = self.param("final_logits_bias", self.bias_init, (1, self.config.decoder_vocab_size))
|
821 |
+
|
822 |
+
def _get_encoder_module(self):
|
823 |
+
return self.model.encoder
|
824 |
+
|
825 |
+
def _get_decoder_module(self):
|
826 |
+
return self.model.decoder
|
827 |
+
|
828 |
+
def __call__(
|
829 |
+
self,
|
830 |
+
input_ids,
|
831 |
+
attention_mask,
|
832 |
+
decoder_input_ids,
|
833 |
+
decoder_attention_mask,
|
834 |
+
position_ids,
|
835 |
+
decoder_position_ids,
|
836 |
+
deterministic: bool = True,
|
837 |
+
):
|
838 |
+
outputs = self.model(
|
839 |
+
input_ids=input_ids,
|
840 |
+
attention_mask=attention_mask,
|
841 |
+
decoder_input_ids=decoder_input_ids,
|
842 |
+
decoder_attention_mask=decoder_attention_mask,
|
843 |
+
position_ids=position_ids,
|
844 |
+
decoder_position_ids=decoder_position_ids,
|
845 |
+
deterministic=deterministic,
|
846 |
+
)
|
847 |
+
|
848 |
+
hidden_states = outputs[0]
|
849 |
+
|
850 |
+
if self.config.tie_word_embeddings:
|
851 |
+
shared_embedding = self.model.variables["params"]["shared"]["embedding"]
|
852 |
+
lm_logits = self.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
|
853 |
+
else:
|
854 |
+
lm_logits = self.lm_head(hidden_states)
|
855 |
+
|
856 |
+
lm_logits += self.final_logits_bias
|
857 |
+
|
858 |
+
return FlaxSeq2SeqLMOutput(
|
859 |
+
logits=lm_logits,
|
860 |
+
decoder_hidden_states=outputs.decoder_hidden_states,
|
861 |
+
decoder_attentions=outputs.decoder_attentions,
|
862 |
+
cross_attentions=outputs.cross_attentions,
|
863 |
+
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
|
864 |
+
encoder_hidden_states=outputs.encoder_hidden_states,
|
865 |
+
encoder_attentions=outputs.encoder_attentions,
|
866 |
+
)
|
867 |
+
|
868 |
+
|
869 |
+
class FlaxBartForConditionalGeneration(FlaxBartPreTrainedModel):
|
870 |
+
module_class = FlaxBartForConditionalGenerationModule
|
871 |
+
dtype: jnp.dtype = jnp.float32
|
872 |
+
|
873 |
+
def decode(
|
874 |
+
self,
|
875 |
+
decoder_input_ids,
|
876 |
+
encoder_outputs,
|
877 |
+
encoder_attention_mask: Optional[jnp.ndarray] = None,
|
878 |
+
decoder_attention_mask: Optional[jnp.ndarray] = None,
|
879 |
+
decoder_position_ids: Optional[jnp.ndarray] = None,
|
880 |
+
past_key_values: dict = None,
|
881 |
+
train: bool = False,
|
882 |
+
params: dict = None,
|
883 |
+
dropout_rng: PRNGKey = None,
|
884 |
+
):
|
885 |
+
r"""
|
886 |
+
Returns:
|
887 |
+
|
888 |
+
Example::
|
889 |
+
|
890 |
+
>>> from transformers import BartTokenizer, FlaxBartForConditionalGeneration
|
891 |
+
|
892 |
+
>>> model = FlaxBartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')
|
893 |
+
>>> tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')
|
894 |
+
|
895 |
+
>>> text = "My friends are cool but they eat too many carbs."
|
896 |
+
>>> inputs = tokenizer(text, max_length=1024, return_tensors='jax')
|
897 |
+
>>> encoder_outputs = model.encode(**inputs)
|
898 |
+
|
899 |
+
>>> decoder_start_token_id = model.config.decoder_start_token_id
|
900 |
+
>>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
|
901 |
+
|
902 |
+
>>> outputs = model.decode(decoder_input_ids, encoder_outputs)
|
903 |
+
>>> logits = outputs.logits
|
904 |
+
"""
|
905 |
+
encoder_hidden_states = encoder_outputs[0]
|
906 |
+
if encoder_attention_mask is None:
|
907 |
+
batch_size, sequence_length = encoder_hidden_states.shape[:2]
|
908 |
+
encoder_attention_mask = jnp.ones((batch_size, sequence_length))
|
909 |
+
|
910 |
+
batch_size, sequence_length = decoder_input_ids.shape
|
911 |
+
if decoder_attention_mask is None:
|
912 |
+
decoder_attention_mask = jnp.ones((batch_size, sequence_length))
|
913 |
+
|
914 |
+
if decoder_position_ids is None:
|
915 |
+
if past_key_values is not None:
|
916 |
+
raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
|
917 |
+
|
918 |
+
decoder_position_ids = jnp.broadcast_to(
|
919 |
+
jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
|
920 |
+
)
|
921 |
+
|
922 |
+
# Handle any PRNG if needed
|
923 |
+
rngs = {}
|
924 |
+
if dropout_rng is not None:
|
925 |
+
rngs["dropout"] = dropout_rng
|
926 |
+
|
927 |
+
inputs = {"params": params or self.params}
|
928 |
+
|
929 |
+
# if past_key_values are passed then cache is already initialized a private flag init_cache has to be
|
930 |
+
# passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
|
931 |
+
# it can be changed by FlaxBartAttention module
|
932 |
+
if past_key_values:
|
933 |
+
inputs["cache"] = past_key_values
|
934 |
+
mutable = ["cache"]
|
935 |
+
else:
|
936 |
+
mutable = False
|
937 |
+
|
938 |
+
def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
|
939 |
+
decoder_module = module._get_decoder_module()
|
940 |
+
outputs = decoder_module(
|
941 |
+
decoder_input_ids,
|
942 |
+
decoder_attention_mask,
|
943 |
+
decoder_position_ids,
|
944 |
+
**kwargs,
|
945 |
+
)
|
946 |
+
hidden_states = outputs[0]
|
947 |
+
|
948 |
+
if self.config.tie_word_embeddings:
|
949 |
+
shared_embedding = module.model.variables["params"]["shared"]["embedding"]
|
950 |
+
lm_logits = module.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
|
951 |
+
else:
|
952 |
+
lm_logits = module.lm_head(hidden_states)
|
953 |
+
|
954 |
+
lm_logits += module.final_logits_bias
|
955 |
+
return lm_logits, outputs
|
956 |
+
|
957 |
+
outputs = self.module.apply(
|
958 |
+
inputs,
|
959 |
+
decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
|
960 |
+
decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
|
961 |
+
decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
|
962 |
+
encoder_hidden_states=encoder_hidden_states,
|
963 |
+
encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
|
964 |
+
deterministic=not train,
|
965 |
+
rngs=rngs,
|
966 |
+
mutable=mutable,
|
967 |
+
method=_decoder_forward,
|
968 |
+
)
|
969 |
+
|
970 |
+
if past_key_values is None:
|
971 |
+
lm_logits, decoder_outputs = outputs
|
972 |
+
else:
|
973 |
+
(lm_logits, decoder_outputs), past = outputs
|
974 |
+
|
975 |
+
outputs = FlaxCausalLMOutputWithCrossAttentions(
|
976 |
+
logits=lm_logits,
|
977 |
+
hidden_states=decoder_outputs.hidden_states,
|
978 |
+
attentions=decoder_outputs.attentions,
|
979 |
+
cross_attentions=decoder_outputs.cross_attentions,
|
980 |
+
)
|
981 |
+
|
982 |
+
# add updated cache to model output
|
983 |
+
if past_key_values is not None:
|
984 |
+
outputs["past_key_values"] = unfreeze(past["cache"])
|
985 |
+
return outputs
|
986 |
+
|
987 |
+
return outputs
|
988 |
+
|
989 |
+
def prepare_inputs_for_generation(
|
990 |
+
self,
|
991 |
+
decoder_input_ids,
|
992 |
+
max_length,
|
993 |
+
attention_mask: Optional[jnp.DeviceArray] = None,
|
994 |
+
decoder_attention_mask: Optional[jnp.DeviceArray] = None,
|
995 |
+
encoder_outputs=None,
|
996 |
+
**kwargs,
|
997 |
+
):
|
998 |
+
# initializing the cache
|
999 |
+
batch_size, seq_length = decoder_input_ids.shape
|
1000 |
+
|
1001 |
+
past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
|
1002 |
+
# Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
|
1003 |
+
# But since the decoder uses a causal mask, those positions are masked anyways.
|
1004 |
+
# Thus we can create a single static attention_mask here, which is more efficient for compilation
|
1005 |
+
extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
|
1006 |
+
if decoder_attention_mask is not None:
|
1007 |
+
position_ids = decoder_attention_mask.cumsum(axis=-1) - 1
|
1008 |
+
extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, decoder_attention_mask, (0, 0))
|
1009 |
+
else:
|
1010 |
+
position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
|
1011 |
+
|
1012 |
+
return {
|
1013 |
+
"past_key_values": past_key_values,
|
1014 |
+
"encoder_outputs": encoder_outputs,
|
1015 |
+
"encoder_attention_mask": attention_mask,
|
1016 |
+
"decoder_attention_mask": extended_attention_mask,
|
1017 |
+
"decoder_position_ids": position_ids,
|
1018 |
+
}
|
1019 |
+
|
1020 |
+
def update_inputs_for_generation(self, model_outputs, model_kwargs):
|
1021 |
+
model_kwargs["past_key_values"] = model_outputs.past_key_values
|
1022 |
+
model_kwargs["decoder_position_ids"] = model_kwargs["decoder_position_ids"][:, -1:] + 1
|
1023 |
+
return model_kwargs
|