Markus28
commited on
Commit
•
87b642a
0
Parent(s):
initial commit
Browse files- configuration_bert.py +95 -0
- modeling_bert.py +760 -0
configuration_bert.py
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
3 |
+
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
4 |
+
#
|
5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
+
# you may not use this file except in compliance with the License.
|
7 |
+
# You may obtain a copy of the License at
|
8 |
+
#
|
9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
10 |
+
#
|
11 |
+
# Unless required by applicable law or agreed to in writing, software
|
12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
+
# See the License for the specific language governing permissions and
|
15 |
+
# limitations under the License.
|
16 |
+
""" BERT model configuration"""
|
17 |
+
from collections import OrderedDict
|
18 |
+
from typing import Mapping
|
19 |
+
|
20 |
+
from transformers import PretrainedConfig
|
21 |
+
|
22 |
+
|
23 |
+
class JinaBertConfig(PretrainedConfig):
|
24 |
+
r"""
|
25 |
+
This is the configuration class to store the configuration of a [`BertModel`] or a [`TFBertModel`]. It is used to
|
26 |
+
instantiate a BERT model according to the specified arguments, defining the model architecture. Instantiating a
|
27 |
+
configuration with the defaults will yield a similar configuration to that of the BERT
|
28 |
+
[google-bert/bert-base-uncased](https://huggingface.co/google-bert/bert-base-uncased) architecture.
|
29 |
+
|
30 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
31 |
+
documentation from [`PretrainedConfig`] for more information.
|
32 |
+
|
33 |
+
|
34 |
+
Args:
|
35 |
+
vocab_size (`int`, *optional*, defaults to 30522):
|
36 |
+
Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
|
37 |
+
`inputs_ids` passed when calling [`BertModel`] or [`TFBertModel`].
|
38 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
39 |
+
Dimensionality of the encoder layers and the pooler layer.
|
40 |
+
num_hidden_layers (`int`, *optional*, defaults to 12):
|
41 |
+
Number of hidden layers in the Transformer encoder.
|
42 |
+
num_attention_heads (`int`, *optional*, defaults to 12):
|
43 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
44 |
+
intermediate_size (`int`, *optional*, defaults to 3072):
|
45 |
+
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
|
46 |
+
hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
|
47 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
48 |
+
`"relu"`, `"silu"` and `"gelu_new"` are supported.
|
49 |
+
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
|
50 |
+
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
51 |
+
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
|
52 |
+
The dropout ratio for the attention probabilities.
|
53 |
+
type_vocab_size (`int`, *optional*, defaults to 2):
|
54 |
+
The vocabulary size of the `token_type_ids` passed when calling [`BertModel`] or [`TFBertModel`].
|
55 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
56 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
57 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
|
58 |
+
The epsilon used by the layer normalization layers.
|
59 |
+
window_size (`tuple`, *optional*, defaults to `(-1, -1)`): If not the default, use local attention
|
60 |
+
"""
|
61 |
+
|
62 |
+
model_type = "bert"
|
63 |
+
|
64 |
+
def __init__(
|
65 |
+
self,
|
66 |
+
vocab_size=30522,
|
67 |
+
hidden_size=768,
|
68 |
+
num_hidden_layers=12,
|
69 |
+
num_attention_heads=12,
|
70 |
+
intermediate_size=3072,
|
71 |
+
hidden_act="gelu",
|
72 |
+
hidden_dropout_prob=0.1,
|
73 |
+
attention_probs_dropout_prob=0.1,
|
74 |
+
type_vocab_size=2,
|
75 |
+
initializer_range=0.02,
|
76 |
+
layer_norm_eps=1e-12,
|
77 |
+
pad_token_id=0,
|
78 |
+
window_size=(-1, -1),
|
79 |
+
**kwargs,
|
80 |
+
):
|
81 |
+
super().__init__(pad_token_id=pad_token_id, **kwargs)
|
82 |
+
|
83 |
+
self.vocab_size = vocab_size
|
84 |
+
self.hidden_size = hidden_size
|
85 |
+
self.num_hidden_layers = num_hidden_layers
|
86 |
+
self.num_attention_heads = num_attention_heads
|
87 |
+
self.hidden_act = hidden_act
|
88 |
+
self.intermediate_size = intermediate_size
|
89 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
90 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
91 |
+
self.type_vocab_size = type_vocab_size
|
92 |
+
self.initializer_range = initializer_range
|
93 |
+
self.layer_norm_eps = layer_norm_eps
|
94 |
+
self.window_size = window_size
|
95 |
+
|
modeling_bert.py
ADDED
@@ -0,0 +1,760 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) 2022, Tri Dao.
|
2 |
+
# This BERT implementation is based on our MLPerf 2.0 and MLPerf 2.1 BERT implementation.
|
3 |
+
# https://github.com/mlcommons/training_results_v2.0/blob/main/HazyResearch/benchmarks/bert/implementations/pytorch/modeling.py
|
4 |
+
# https://github.com/mlcommons/training_results_v2.1/blob/main/Azure-HazyResearch/benchmarks/bert/implementations/ND96amsr_A100_v4/modeling.py
|
5 |
+
|
6 |
+
# Inspired by https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py
|
7 |
+
|
8 |
+
import logging
|
9 |
+
import re
|
10 |
+
from collections import OrderedDict
|
11 |
+
from collections.abc import Sequence
|
12 |
+
from functools import partial
|
13 |
+
|
14 |
+
import torch
|
15 |
+
import torch.nn as nn
|
16 |
+
import torch.nn.functional as F
|
17 |
+
from einops import rearrange
|
18 |
+
from transformers import PretrainedConfig
|
19 |
+
from configuration_bert import JinaBertConfig
|
20 |
+
from transformers.models.bert.modeling_bert import (
|
21 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
22 |
+
BertForPreTrainingOutput,
|
23 |
+
)
|
24 |
+
|
25 |
+
from flash_attn.bert_padding import (
|
26 |
+
index_first_axis,
|
27 |
+
index_first_axis_residual,
|
28 |
+
pad_input,
|
29 |
+
unpad_input,
|
30 |
+
)
|
31 |
+
from flash_attn.modules.block import Block
|
32 |
+
from flash_attn.modules.embedding import BertEmbeddings
|
33 |
+
from flash_attn.modules.mha import MHA
|
34 |
+
from flash_attn.modules.mlp import FusedMLP, Mlp
|
35 |
+
from flash_attn.utils.pretrained import state_dict_from_pretrained
|
36 |
+
|
37 |
+
try:
|
38 |
+
from flash_attn.ops.fused_dense import FusedDense
|
39 |
+
except ImportError:
|
40 |
+
FusedDense = None
|
41 |
+
|
42 |
+
try:
|
43 |
+
from flash_attn.ops.triton.layer_norm import layer_norm_fn
|
44 |
+
except ImportError:
|
45 |
+
layer_norm_fn = None
|
46 |
+
|
47 |
+
|
48 |
+
try:
|
49 |
+
from flash_attn.losses.cross_entropy import CrossEntropyLoss
|
50 |
+
except ImportError:
|
51 |
+
CrossEntropyLoss = None
|
52 |
+
|
53 |
+
|
54 |
+
logger = logging.getLogger(__name__)
|
55 |
+
|
56 |
+
|
57 |
+
def create_mixer_cls(config, cross_attn=False, return_residual=False):
|
58 |
+
fused_bias_fc = getattr(config, "fused_bias_fc", False)
|
59 |
+
window_size = getattr(config, "window_size", (-1, -1))
|
60 |
+
mixer_cls = partial(
|
61 |
+
MHA,
|
62 |
+
num_heads=config.num_attention_heads,
|
63 |
+
cross_attn=cross_attn,
|
64 |
+
dropout=config.attention_probs_dropout_prob,
|
65 |
+
causal=False,
|
66 |
+
fused_bias_fc=fused_bias_fc,
|
67 |
+
use_flash_attn=True,
|
68 |
+
return_residual=return_residual,
|
69 |
+
use_alibi=True,
|
70 |
+
window_size=window_size,
|
71 |
+
)
|
72 |
+
return mixer_cls
|
73 |
+
|
74 |
+
|
75 |
+
def create_mlp_cls(config, layer_idx=None, return_residual=False):
|
76 |
+
inner_dim = config.intermediate_size
|
77 |
+
fused_mlp = getattr(config, "fused_mlp", False)
|
78 |
+
if fused_mlp:
|
79 |
+
assert config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"], (
|
80 |
+
"fused_mlp only " "supports approximate gelu"
|
81 |
+
)
|
82 |
+
if not fused_mlp:
|
83 |
+
approximate = (
|
84 |
+
"tanh"
|
85 |
+
if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
|
86 |
+
else "none"
|
87 |
+
)
|
88 |
+
mlp_cls = partial(
|
89 |
+
Mlp,
|
90 |
+
hidden_features=inner_dim,
|
91 |
+
activation=partial(F.gelu, approximate=approximate),
|
92 |
+
return_residual=return_residual,
|
93 |
+
)
|
94 |
+
else:
|
95 |
+
if FusedMLP is None:
|
96 |
+
raise ImportError("fused_dense is not installed")
|
97 |
+
mlp_checkpoint_lvl = getattr(config, "mlp_checkpoint_lvl", 0)
|
98 |
+
# mlp_checkpoint_lvl could be a list, which contains the checkpoint_lvl for each layer
|
99 |
+
if isinstance(mlp_checkpoint_lvl, Sequence):
|
100 |
+
assert layer_idx is not None
|
101 |
+
mlp_checkpoint_lvl = mlp_checkpoint_lvl[layer_idx]
|
102 |
+
mlp_cls = partial(
|
103 |
+
FusedMLP,
|
104 |
+
hidden_features=inner_dim,
|
105 |
+
checkpoint_lvl=mlp_checkpoint_lvl,
|
106 |
+
return_residual=return_residual,
|
107 |
+
)
|
108 |
+
return mlp_cls
|
109 |
+
|
110 |
+
|
111 |
+
def create_block(config, layer_idx=None):
|
112 |
+
last_layer_subset = getattr(config, "last_layer_subset", False)
|
113 |
+
cross_attn = last_layer_subset and layer_idx == config.num_hidden_layers - 1
|
114 |
+
# TD [2022-12-19]: For cross attention (last layer), we actually want to return the
|
115 |
+
# residual x_kv, not residual x. But it's annoying to change the API (and it only affects
|
116 |
+
# one layer) so we just choose not to return residual in this case.
|
117 |
+
return_residual = not cross_attn
|
118 |
+
mixer_cls = create_mixer_cls(config, cross_attn, return_residual=return_residual)
|
119 |
+
mlp_cls = create_mlp_cls(config, layer_idx, return_residual=return_residual)
|
120 |
+
norm_cls = partial(nn.LayerNorm, eps=config.layer_norm_eps)
|
121 |
+
block = Block(
|
122 |
+
config.hidden_size,
|
123 |
+
mixer_cls,
|
124 |
+
mlp_cls,
|
125 |
+
norm_cls=norm_cls,
|
126 |
+
prenorm=False,
|
127 |
+
resid_dropout1=config.hidden_dropout_prob,
|
128 |
+
resid_dropout2=config.hidden_dropout_prob,
|
129 |
+
fused_dropout_add_ln=getattr(config, "fused_dropout_add_ln", False),
|
130 |
+
return_residual=return_residual,
|
131 |
+
)
|
132 |
+
return block
|
133 |
+
|
134 |
+
|
135 |
+
# https://github.com/huggingface/transformers/blob/7032e0203262ebb2ebf55da8d2e01f873973e835/src/transformers/models/bert/modeling_bert.py#L748
|
136 |
+
def _init_weights(module, initializer_range=0.02):
|
137 |
+
if isinstance(module, nn.Linear):
|
138 |
+
nn.init.normal_(module.weight, std=initializer_range)
|
139 |
+
if module.bias is not None:
|
140 |
+
nn.init.zeros_(module.bias)
|
141 |
+
elif isinstance(module, nn.Embedding):
|
142 |
+
nn.init.normal_(module.weight, std=initializer_range)
|
143 |
+
if module.padding_idx is not None:
|
144 |
+
nn.init.zeros_(module.weight[module.padding_idx])
|
145 |
+
|
146 |
+
|
147 |
+
class BertEncoder(nn.Module):
|
148 |
+
def __init__(self, config: JinaBertConfig):
|
149 |
+
super().__init__()
|
150 |
+
self.use_flash_attn = getattr(config, "use_flash_attn", False)
|
151 |
+
self.layers = nn.ModuleList(
|
152 |
+
[create_block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
|
153 |
+
)
|
154 |
+
|
155 |
+
def forward(self, hidden_states, key_padding_mask=None, subset_mask=None):
|
156 |
+
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
157 |
+
This means that we only compute the last layer output for these tokens.
|
158 |
+
subset_mask: (batch, seqlen), dtype=torch.bool
|
159 |
+
"""
|
160 |
+
if key_padding_mask is None or not self.use_flash_attn:
|
161 |
+
mixer_kwargs = (
|
162 |
+
{"key_padding_mask": key_padding_mask} if key_padding_mask is not None else None
|
163 |
+
)
|
164 |
+
for layer in self.layers:
|
165 |
+
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
166 |
+
print(hidden_states)
|
167 |
+
if subset_mask is not None:
|
168 |
+
hidden_states = hidden_states[subset_mask]
|
169 |
+
else:
|
170 |
+
batch, seqlen = hidden_states.shape[:2]
|
171 |
+
hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
|
172 |
+
hidden_states, key_padding_mask
|
173 |
+
)
|
174 |
+
mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch}
|
175 |
+
if subset_mask is None:
|
176 |
+
for layer in self.layers:
|
177 |
+
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
178 |
+
hidden_states = pad_input(hidden_states, indices, batch, seqlen)
|
179 |
+
else:
|
180 |
+
for layer in self.layers[:-1]:
|
181 |
+
hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
|
182 |
+
if key_padding_mask is not None:
|
183 |
+
subset_idx = torch.nonzero(
|
184 |
+
subset_mask[key_padding_mask], as_tuple=False
|
185 |
+
).flatten()
|
186 |
+
subset_seqlens = (subset_mask & key_padding_mask).sum(dim=-1, dtype=torch.int32)
|
187 |
+
subset_cu_seqlens = F.pad(
|
188 |
+
torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32), (1, 0)
|
189 |
+
)
|
190 |
+
else:
|
191 |
+
subset_idx = torch.nonzero(subset_mask, as_tuple=False).flatten()
|
192 |
+
subset_seqlens = subset_mask.sum(dim=-1, dtype=torch.int32)
|
193 |
+
subset_cu_seqlens = F.pad(
|
194 |
+
torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32), (1, 0)
|
195 |
+
)
|
196 |
+
hidden_states_subset, hidden_states = index_first_axis_residual(
|
197 |
+
hidden_states, subset_idx
|
198 |
+
)
|
199 |
+
# It's ok to set max_seqlen_q to be much larger
|
200 |
+
mixer_kwargs = {
|
201 |
+
"x_kv": hidden_states,
|
202 |
+
"cu_seqlens": subset_cu_seqlens,
|
203 |
+
"max_seqlen": max_seqlen_in_batch,
|
204 |
+
"cu_seqlens_k": cu_seqlens,
|
205 |
+
"max_seqlen_k": max_seqlen_in_batch,
|
206 |
+
}
|
207 |
+
hidden_states = self.layers[-1](hidden_states_subset, mixer_kwargs=mixer_kwargs)
|
208 |
+
return hidden_states
|
209 |
+
|
210 |
+
|
211 |
+
class BertPooler(nn.Module):
|
212 |
+
def __init__(self, config):
|
213 |
+
super().__init__()
|
214 |
+
fused_bias_fc = getattr(config, "fused_bias_fc", False)
|
215 |
+
if fused_bias_fc and FusedDense is None:
|
216 |
+
raise ImportError("fused_dense is not installed")
|
217 |
+
linear_cls = nn.Linear if not fused_bias_fc else FusedDense
|
218 |
+
self.dense = linear_cls(config.hidden_size, config.hidden_size)
|
219 |
+
self.activation = nn.Tanh()
|
220 |
+
|
221 |
+
def forward(self, hidden_states, pool=True):
|
222 |
+
# We "pool" the model by simply taking the hidden state corresponding
|
223 |
+
# to the first token.
|
224 |
+
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
225 |
+
pooled_output = self.dense(first_token_tensor)
|
226 |
+
pooled_output = self.activation(pooled_output)
|
227 |
+
return pooled_output
|
228 |
+
|
229 |
+
|
230 |
+
class BertPredictionHeadTransform(nn.Module):
|
231 |
+
def __init__(self, config):
|
232 |
+
super().__init__()
|
233 |
+
fused_bias_fc = getattr(config, "fused_bias_fc", False)
|
234 |
+
if fused_bias_fc and FusedDense is None:
|
235 |
+
raise ImportError("fused_dense is not installed")
|
236 |
+
self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
|
237 |
+
if self.fused_dropout_add_ln and layer_norm_fn is None:
|
238 |
+
raise ImportError("Triton is not installed")
|
239 |
+
linear_cls = nn.Linear if not fused_bias_fc else FusedDense
|
240 |
+
self.dense = linear_cls(config.hidden_size, config.hidden_size)
|
241 |
+
approximate = (
|
242 |
+
"tanh"
|
243 |
+
if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
|
244 |
+
else "none"
|
245 |
+
)
|
246 |
+
self.transform_act_fn = nn.GELU(approximate=approximate)
|
247 |
+
self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
248 |
+
|
249 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
250 |
+
hidden_states = self.dense(hidden_states)
|
251 |
+
hidden_states = self.transform_act_fn(hidden_states)
|
252 |
+
if not self.fused_dropout_add_ln:
|
253 |
+
hidden_states = self.layer_norm(hidden_states)
|
254 |
+
else:
|
255 |
+
hidden_states = layer_norm_fn(
|
256 |
+
hidden_states, self.layer_norm.weight, self.layer_norm.bias, eps=self.layer_norm.eps
|
257 |
+
)
|
258 |
+
return hidden_states
|
259 |
+
|
260 |
+
|
261 |
+
class BertLMPredictionHead(nn.Module):
|
262 |
+
def __init__(self, config):
|
263 |
+
super().__init__()
|
264 |
+
fused_bias_fc = getattr(config, "fused_bias_fc", False)
|
265 |
+
if fused_bias_fc and FusedDense is None:
|
266 |
+
raise ImportError("fused_dense is not installed")
|
267 |
+
linear_cls = nn.Linear if not fused_bias_fc else FusedDense
|
268 |
+
|
269 |
+
self.transform = BertPredictionHeadTransform(config)
|
270 |
+
|
271 |
+
# The output weights are the same as the input embeddings, but there is
|
272 |
+
# an output-only bias for each token.
|
273 |
+
self.decoder = linear_cls(config.hidden_size, config.vocab_size, bias=True)
|
274 |
+
|
275 |
+
def forward(self, hidden_states):
|
276 |
+
hidden_states = self.transform(hidden_states)
|
277 |
+
hidden_states = self.decoder(hidden_states)
|
278 |
+
return hidden_states
|
279 |
+
|
280 |
+
|
281 |
+
class BertPreTrainingHeads(nn.Module):
|
282 |
+
def __init__(self, config):
|
283 |
+
super().__init__()
|
284 |
+
self.predictions = BertLMPredictionHead(config)
|
285 |
+
self.seq_relationship = nn.Linear(config.hidden_size, 2)
|
286 |
+
|
287 |
+
def forward(self, sequence_output, pooled_output):
|
288 |
+
prediction_scores = self.predictions(sequence_output)
|
289 |
+
seq_relationship_score = self.seq_relationship(pooled_output)
|
290 |
+
return prediction_scores, seq_relationship_score
|
291 |
+
|
292 |
+
|
293 |
+
class BertPreTrainedModel(nn.Module):
|
294 |
+
"""An abstract class to handle weights initialization and
|
295 |
+
a simple interface for dowloading and loading pretrained models.
|
296 |
+
"""
|
297 |
+
|
298 |
+
def __init__(self, config, *inputs, **kwargs):
|
299 |
+
super().__init__()
|
300 |
+
if not isinstance(config, JinaBertConfig):
|
301 |
+
raise ValueError(
|
302 |
+
"Parameter config in `{}(config)` should be an instance of class `BertConfig`. "
|
303 |
+
"To create a model from a Google pretrained model use "
|
304 |
+
"`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
|
305 |
+
self.__class__.__name__, self.__class__.__name__
|
306 |
+
)
|
307 |
+
)
|
308 |
+
self.config = config
|
309 |
+
|
310 |
+
@classmethod
|
311 |
+
def from_pretrained(cls, model_name, config, *inputs, **kwargs):
|
312 |
+
"""
|
313 |
+
Instantiate a BertPreTrainedModel from a pre-trained model file or a pytorch state dict.
|
314 |
+
Download and cache the pre-trained model file if needed.
|
315 |
+
|
316 |
+
Params:
|
317 |
+
pretrained_model_name_or_path: either:
|
318 |
+
- a path or url to a pretrained model archive containing:
|
319 |
+
. `bert_config.json` a configuration file for the model
|
320 |
+
. `pytorch_model.bin` a PyTorch dump of a BertForPretraining instance
|
321 |
+
- a path or url to a pretrained model archive containing:
|
322 |
+
. `bert_config.json` a configuration file for the model
|
323 |
+
. `model.chkpt` a TensorFlow checkpoint
|
324 |
+
*inputs, **kwargs: additional input for the specific Bert class
|
325 |
+
(ex: num_labels for BertForSequenceClassification)
|
326 |
+
"""
|
327 |
+
# Instantiate model.
|
328 |
+
model = cls(config, *inputs, **kwargs)
|
329 |
+
load_return = model.load_state_dict(
|
330 |
+
remap_state_dict(state_dict_from_pretrained(model_name), config), strict=False
|
331 |
+
)
|
332 |
+
logger.info(load_return)
|
333 |
+
return model
|
334 |
+
|
335 |
+
|
336 |
+
class BertModel(BertPreTrainedModel):
|
337 |
+
def __init__(self, config: JinaBertConfig, add_pooling_layer=True):
|
338 |
+
super().__init__(config)
|
339 |
+
self.pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
340 |
+
if config.vocab_size % self.pad_vocab_size_multiple != 0:
|
341 |
+
config.vocab_size += self.pad_vocab_size_multiple - (
|
342 |
+
config.vocab_size % self.pad_vocab_size_multiple
|
343 |
+
)
|
344 |
+
self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
|
345 |
+
if self.fused_dropout_add_ln and layer_norm_fn is None:
|
346 |
+
raise ImportError("Triton is not installed")
|
347 |
+
assert config.hidden_act in ["gelu", "gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
|
348 |
+
|
349 |
+
self.embeddings = BertEmbeddings(
|
350 |
+
config.hidden_size,
|
351 |
+
config.vocab_size,
|
352 |
+
-1, # No position embeddings
|
353 |
+
config.type_vocab_size,
|
354 |
+
padding_idx=config.pad_token_id,
|
355 |
+
)
|
356 |
+
self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
|
357 |
+
self.emb_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
358 |
+
self.encoder = BertEncoder(config)
|
359 |
+
self.pooler = BertPooler(config) if add_pooling_layer else None
|
360 |
+
|
361 |
+
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
362 |
+
|
363 |
+
def forward(
|
364 |
+
self,
|
365 |
+
input_ids,
|
366 |
+
position_ids=None,
|
367 |
+
token_type_ids=None,
|
368 |
+
attention_mask=None,
|
369 |
+
masked_tokens_mask=None,
|
370 |
+
):
|
371 |
+
"""If masked_tokens_mask is not None (i.e. last_layer_subset == True in BertForPreTraining),
|
372 |
+
we only want the output for the masked tokens. This means that we only compute the last
|
373 |
+
layer output for these tokens.
|
374 |
+
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
375 |
+
"""
|
376 |
+
hidden_states = self.embeddings(
|
377 |
+
input_ids, position_ids=position_ids, token_type_ids=token_type_ids
|
378 |
+
)
|
379 |
+
# TD [2022-12:18]: Don't need to force residual in fp32
|
380 |
+
# BERT puts embedding LayerNorm before embedding dropout.
|
381 |
+
if not self.fused_dropout_add_ln:
|
382 |
+
hidden_states = self.emb_ln(hidden_states)
|
383 |
+
else:
|
384 |
+
hidden_states = layer_norm_fn(
|
385 |
+
hidden_states, self.emb_ln.weight, self.emb_ln.bias, eps=self.emb_ln.eps
|
386 |
+
)
|
387 |
+
hidden_states = self.emb_drop(hidden_states)
|
388 |
+
|
389 |
+
if masked_tokens_mask is not None:
|
390 |
+
batch_size, seqlen = input_ids.shape[:2]
|
391 |
+
# We also need the first column for the CLS token
|
392 |
+
first_col_mask = torch.zeros(
|
393 |
+
batch_size, seqlen, dtype=torch.bool, device=input_ids.device
|
394 |
+
)
|
395 |
+
first_col_mask[:, 0] = True
|
396 |
+
subset_mask = masked_tokens_mask | first_col_mask
|
397 |
+
else:
|
398 |
+
subset_mask = None
|
399 |
+
|
400 |
+
sequence_output = self.encoder(
|
401 |
+
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
|
402 |
+
)
|
403 |
+
|
404 |
+
if masked_tokens_mask is None:
|
405 |
+
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
|
406 |
+
else:
|
407 |
+
# TD [2022-03-01]: the indexing here is very tricky.
|
408 |
+
if attention_mask is not None:
|
409 |
+
subset_idx = subset_mask[attention_mask]
|
410 |
+
pool_input = sequence_output[first_col_mask[attention_mask][subset_idx]]
|
411 |
+
sequence_output = sequence_output[masked_tokens_mask[attention_mask][subset_idx]]
|
412 |
+
else:
|
413 |
+
pool_input = sequence_output[first_col_mask[subset_mask]]
|
414 |
+
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
415 |
+
pooled_output = self.pooler(pool_input, pool=False) if self.pooler is not None else None
|
416 |
+
|
417 |
+
return BaseModelOutputWithPoolingAndCrossAttentions(
|
418 |
+
last_hidden_state=sequence_output,
|
419 |
+
pooler_output=pooled_output,
|
420 |
+
)
|
421 |
+
|
422 |
+
|
423 |
+
class BertForPreTraining(BertPreTrainedModel):
|
424 |
+
def __init__(self, config: JinaBertConfig):
|
425 |
+
super().__init__(config)
|
426 |
+
# If dense_seq_output, we only need to pass the hidden states for the masked out tokens
|
427 |
+
# (around 15%) to the classifier heads.
|
428 |
+
self.dense_seq_output = getattr(config, "dense_seq_output", False)
|
429 |
+
# If last_layer_subset, we only need the compute the last layer for a subset of tokens
|
430 |
+
# (e.g., the tokens we need to compute the masked LM loss and the next-sentence prediction).
|
431 |
+
self.last_layer_subset = getattr(config, "last_layer_subset", False)
|
432 |
+
if self.last_layer_subset:
|
433 |
+
assert self.dense_seq_output, "last_layer_subset requires dense_seq_output"
|
434 |
+
use_xentropy = getattr(config, "use_xentropy", False)
|
435 |
+
if use_xentropy and CrossEntropyLoss is None:
|
436 |
+
raise ImportError("xentropy_cuda is not installed")
|
437 |
+
loss_cls = (
|
438 |
+
nn.CrossEntropyLoss
|
439 |
+
if not use_xentropy
|
440 |
+
else partial(CrossEntropyLoss, inplace_backward=True)
|
441 |
+
)
|
442 |
+
|
443 |
+
self.bert = BertModel(config)
|
444 |
+
self.cls = BertPreTrainingHeads(config)
|
445 |
+
self.mlm_loss = loss_cls(ignore_index=0)
|
446 |
+
self.nsp_loss = loss_cls(ignore_index=-1)
|
447 |
+
|
448 |
+
# Initialize weights and apply final processing
|
449 |
+
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
450 |
+
self.tie_weights()
|
451 |
+
|
452 |
+
def tie_weights(self):
|
453 |
+
self.cls.predictions.decoder.weight = self.bert.embeddings.word_embeddings.weight
|
454 |
+
|
455 |
+
def forward(
|
456 |
+
self,
|
457 |
+
input_ids,
|
458 |
+
position_ids=None,
|
459 |
+
token_type_ids=None,
|
460 |
+
attention_mask=None,
|
461 |
+
labels=None,
|
462 |
+
next_sentence_label=None,
|
463 |
+
):
|
464 |
+
"""
|
465 |
+
If labels are provided, they must be 0 for masked out tokens (as specified in the attention
|
466 |
+
mask).
|
467 |
+
Outputs:
|
468 |
+
if `labels` and `next_sentence_label` are not `None`:
|
469 |
+
Outputs the total_loss which is the sum of the masked language modeling loss and the next
|
470 |
+
sentence classification loss.
|
471 |
+
if `labels` or `next_sentence_label` is `None`:
|
472 |
+
Outputs a tuple comprising
|
473 |
+
- the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
|
474 |
+
- the next sentence classification logits of shape [batch_size, 2].
|
475 |
+
|
476 |
+
"""
|
477 |
+
masked_tokens_mask = labels > 0 if (self.last_layer_subset and labels is not None) else None
|
478 |
+
outputs = self.bert(
|
479 |
+
input_ids,
|
480 |
+
position_ids=position_ids,
|
481 |
+
token_type_ids=token_type_ids,
|
482 |
+
attention_mask=attention_mask.bool() if attention_mask is not None else None,
|
483 |
+
masked_tokens_mask=masked_tokens_mask,
|
484 |
+
)
|
485 |
+
sequence_output, pooled_output = outputs.last_hidden_state, outputs.pooler_output
|
486 |
+
if self.dense_seq_output and labels is not None:
|
487 |
+
masked_token_idx = torch.nonzero(labels.flatten() > 0, as_tuple=False).flatten()
|
488 |
+
if not self.last_layer_subset:
|
489 |
+
sequence_output = index_first_axis(
|
490 |
+
rearrange(sequence_output, "b s d -> (b s) d"), masked_token_idx
|
491 |
+
)
|
492 |
+
prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
|
493 |
+
|
494 |
+
total_loss = None
|
495 |
+
if labels is not None and next_sentence_label is not None:
|
496 |
+
if (
|
497 |
+
self.dense_seq_output and labels is not None
|
498 |
+
): # prediction_scores are already flattened
|
499 |
+
masked_lm_loss = self.mlm_loss(
|
500 |
+
prediction_scores, labels.flatten()[masked_token_idx]
|
501 |
+
)
|
502 |
+
else:
|
503 |
+
masked_lm_loss = self.mlm_loss(
|
504 |
+
rearrange(prediction_scores, "... v -> (...) v"),
|
505 |
+
rearrange(labels, "... -> (...)"),
|
506 |
+
)
|
507 |
+
next_sentence_loss = self.nsp_loss(
|
508 |
+
rearrange(seq_relationship_score, "... t -> (...) t"),
|
509 |
+
rearrange(next_sentence_label, "... -> (...)"),
|
510 |
+
)
|
511 |
+
total_loss = masked_lm_loss.float() + next_sentence_loss.float()
|
512 |
+
|
513 |
+
return BertForPreTrainingOutput(
|
514 |
+
loss=total_loss,
|
515 |
+
prediction_logits=prediction_scores,
|
516 |
+
seq_relationship_logits=seq_relationship_score,
|
517 |
+
)
|
518 |
+
|
519 |
+
|
520 |
+
def remap_state_dict(state_dict, config: PretrainedConfig):
|
521 |
+
"""
|
522 |
+
Map the state_dict of a Huggingface BERT model to be flash_attn compatible.
|
523 |
+
"""
|
524 |
+
|
525 |
+
# LayerNorm
|
526 |
+
def key_mapping_ln_gamma_beta(key):
|
527 |
+
key = re.sub(r"LayerNorm.gamma$", "LayerNorm.weight", key)
|
528 |
+
key = re.sub(r"LayerNorm.beta$", "LayerNorm.bias", key)
|
529 |
+
return key
|
530 |
+
|
531 |
+
state_dict = OrderedDict((key_mapping_ln_gamma_beta(k), v) for k, v in state_dict.items())
|
532 |
+
|
533 |
+
# Layers
|
534 |
+
def key_mapping_layers(key):
|
535 |
+
return re.sub(r"^bert.encoder.layer.", "bert.encoder.layers.", key)
|
536 |
+
|
537 |
+
state_dict = OrderedDict((key_mapping_layers(k), v) for k, v in state_dict.items())
|
538 |
+
|
539 |
+
# LayerNorm
|
540 |
+
def key_mapping_ln(key):
|
541 |
+
key = re.sub(r"^bert.embeddings.LayerNorm.", "bert.emb_ln.", key)
|
542 |
+
key = re.sub(
|
543 |
+
r"^bert.encoder.layers.(\d+).attention.output.LayerNorm.(weight|bias)",
|
544 |
+
r"bert.encoder.layers.\1.norm1.\2",
|
545 |
+
key,
|
546 |
+
)
|
547 |
+
key = re.sub(
|
548 |
+
r"^bert.encoder.layers.(\d+).output.LayerNorm.(weight|bias)",
|
549 |
+
r"bert.encoder.layers.\1.norm2.\2",
|
550 |
+
key,
|
551 |
+
)
|
552 |
+
key = re.sub(
|
553 |
+
r"^cls.predictions.transform.LayerNorm.(weight|bias)",
|
554 |
+
r"cls.predictions.transform.layer_norm.\1",
|
555 |
+
key,
|
556 |
+
)
|
557 |
+
return key
|
558 |
+
|
559 |
+
state_dict = OrderedDict((key_mapping_ln(k), v) for k, v in state_dict.items())
|
560 |
+
|
561 |
+
# MLP
|
562 |
+
def key_mapping_mlp(key):
|
563 |
+
key = re.sub(
|
564 |
+
r"^bert.encoder.layers.(\d+).intermediate.dense.(weight|bias)",
|
565 |
+
r"bert.encoder.layers.\1.mlp.fc1.\2",
|
566 |
+
key,
|
567 |
+
)
|
568 |
+
key = re.sub(
|
569 |
+
r"^bert.encoder.layers.(\d+).output.dense.(weight|bias)",
|
570 |
+
r"bert.encoder.layers.\1.mlp.fc2.\2",
|
571 |
+
key,
|
572 |
+
)
|
573 |
+
return key
|
574 |
+
|
575 |
+
state_dict = OrderedDict((key_mapping_mlp(k), v) for k, v in state_dict.items())
|
576 |
+
|
577 |
+
# Attention
|
578 |
+
last_layer_subset = getattr(config, "last_layer_subset", False)
|
579 |
+
for d in range(config.num_hidden_layers):
|
580 |
+
Wq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.weight")
|
581 |
+
Wk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.weight")
|
582 |
+
Wv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.weight")
|
583 |
+
bq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.bias")
|
584 |
+
bk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.bias")
|
585 |
+
bv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.bias")
|
586 |
+
if not (last_layer_subset and d == config.num_hidden_layers - 1):
|
587 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.weight"] = torch.cat(
|
588 |
+
[Wq, Wk, Wv], dim=0
|
589 |
+
)
|
590 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.bias"] = torch.cat([bq, bk, bv], dim=0)
|
591 |
+
else:
|
592 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wq.weight"] = Wq
|
593 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.weight"] = torch.cat([Wk, Wv], dim=0)
|
594 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wq.bias"] = bq
|
595 |
+
state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.bias"] = torch.cat([bk, bv], dim=0)
|
596 |
+
|
597 |
+
def key_mapping_attn(key):
|
598 |
+
return re.sub(
|
599 |
+
r"^bert.encoder.layers.(\d+).attention.output.dense.(weight|bias)",
|
600 |
+
r"bert.encoder.layers.\1.mixer.out_proj.\2",
|
601 |
+
key,
|
602 |
+
)
|
603 |
+
|
604 |
+
state_dict = OrderedDict((key_mapping_attn(k), v) for k, v in state_dict.items())
|
605 |
+
|
606 |
+
def key_mapping_decoder_bias(key):
|
607 |
+
return re.sub(r"^cls.predictions.bias", "cls.predictions.decoder.bias", key)
|
608 |
+
|
609 |
+
state_dict = OrderedDict((key_mapping_decoder_bias(k), v) for k, v in state_dict.items())
|
610 |
+
|
611 |
+
# Word embedding
|
612 |
+
pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
613 |
+
if pad_vocab_size_multiple > 1:
|
614 |
+
word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
|
615 |
+
state_dict["bert.embeddings.word_embeddings.weight"] = F.pad(
|
616 |
+
word_embeddings, (0, 0, 0, config.vocab_size - word_embeddings.shape[0])
|
617 |
+
)
|
618 |
+
decoder_weight = state_dict["cls.predictions.decoder.weight"]
|
619 |
+
state_dict["cls.predictions.decoder.weight"] = F.pad(
|
620 |
+
decoder_weight, (0, 0, 0, config.vocab_size - decoder_weight.shape[0])
|
621 |
+
)
|
622 |
+
# If the vocab was padded, we want to set the decoder bias for those padded indices to be
|
623 |
+
# strongly negative (i.e. the decoder shouldn't predict those indices).
|
624 |
+
# TD [2022-05-09]: I don't think it affects the MLPerf training.
|
625 |
+
decoder_bias = state_dict["cls.predictions.decoder.bias"]
|
626 |
+
state_dict["cls.predictions.decoder.bias"] = F.pad(
|
627 |
+
decoder_bias, (0, config.vocab_size - decoder_bias.shape[0]), value=-100.0
|
628 |
+
)
|
629 |
+
|
630 |
+
return state_dict
|
631 |
+
|
632 |
+
|
633 |
+
def inv_remap_state_dict(state_dict, config: PretrainedConfig):
|
634 |
+
"""
|
635 |
+
Map the state_dict of a flash_attn model to be Huggingface BERT compatible.
|
636 |
+
|
637 |
+
This function is meant to be the inverse of remap_state_dict.
|
638 |
+
"""
|
639 |
+
# Word embedding
|
640 |
+
pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
641 |
+
if pad_vocab_size_multiple > 1:
|
642 |
+
word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
|
643 |
+
decoder_weight = state_dict["cls.predictions.decoder.weight"]
|
644 |
+
decoder_bias = state_dict["cls.predictions.decoder.bias"]
|
645 |
+
# unpad embeddings
|
646 |
+
state_dict["bert.embeddings.word_embeddings.weight"] = word_embeddings[
|
647 |
+
: config.orig_vocab_size, :
|
648 |
+
]
|
649 |
+
state_dict["cls.predictions.decoder.weight"] = decoder_weight[: config.orig_vocab_size, :]
|
650 |
+
state_dict["cls.predictions.decoder.bias"] = decoder_bias[: config.orig_vocab_size]
|
651 |
+
|
652 |
+
for d in range(config.num_hidden_layers):
|
653 |
+
last_layer_subset = getattr(config, "last_layer_subset", False)
|
654 |
+
if not last_layer_subset or d != (config.num_hidden_layers - 1):
|
655 |
+
Wqkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.weight")
|
656 |
+
Wqkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.bias")
|
657 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = Wqkv_weights[
|
658 |
+
: Wqkv_weights.shape[0] // 3, :
|
659 |
+
]
|
660 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = Wqkv_weights[
|
661 |
+
Wqkv_weights.shape[0] // 3 : 2 * Wqkv_weights.shape[0] // 3, :
|
662 |
+
]
|
663 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = Wqkv_weights[
|
664 |
+
2 * Wqkv_weights.shape[0] // 3 :, :
|
665 |
+
]
|
666 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wqkv_biases[
|
667 |
+
: Wqkv_biases.shape[0] // 3
|
668 |
+
]
|
669 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wqkv_biases[
|
670 |
+
Wqkv_biases.shape[0] // 3 : 2 * Wqkv_biases.shape[0] // 3
|
671 |
+
]
|
672 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = Wqkv_biases[
|
673 |
+
2 * Wqkv_biases.shape[0] // 3 :
|
674 |
+
]
|
675 |
+
else:
|
676 |
+
Wq_weight = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.weight")
|
677 |
+
Wkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.weight")
|
678 |
+
Wq_bias = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.bias")
|
679 |
+
Wkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.bias")
|
680 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = Wq_weight
|
681 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = Wkv_weights[
|
682 |
+
: Wkv_weights.shape[0] // 2, :
|
683 |
+
]
|
684 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = Wkv_weights[
|
685 |
+
Wkv_weights.shape[0] // 2 :, :
|
686 |
+
]
|
687 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wq_bias
|
688 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wkv_biases[
|
689 |
+
: Wkv_biases.shape[0] // 2
|
690 |
+
]
|
691 |
+
state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = Wkv_biases[
|
692 |
+
Wkv_biases.shape[0] // 2 :
|
693 |
+
]
|
694 |
+
|
695 |
+
def inv_key_mapping_ln(key):
|
696 |
+
key = re.sub(r"bert.emb_ln.", "bert.embeddings.LayerNorm.", key)
|
697 |
+
key = re.sub(
|
698 |
+
r"bert.encoder.layers.(\d+).norm1.(weight|bias)",
|
699 |
+
r"bert.encoder.layers.\1.attention.output.LayerNorm.\2",
|
700 |
+
key,
|
701 |
+
)
|
702 |
+
key = re.sub(
|
703 |
+
r"bert.encoder.layers.(\d+).norm2.(weight|bias)",
|
704 |
+
r"bert.encoder.layers.\1.output.LayerNorm.\2",
|
705 |
+
key,
|
706 |
+
)
|
707 |
+
key = re.sub(
|
708 |
+
r"cls.predictions.transform.layer_norm.(weight|bias)",
|
709 |
+
r"cls.predictions.transform.LayerNorm.\1",
|
710 |
+
key,
|
711 |
+
)
|
712 |
+
return key
|
713 |
+
|
714 |
+
def inv_key_mapping_ln_gamma_beta(key):
|
715 |
+
key = re.sub(r"LayerNorm.weight$", "LayerNorm.gamma", key)
|
716 |
+
key = re.sub(r"LayerNorm.bias$", "LayerNorm.beta", key)
|
717 |
+
return key
|
718 |
+
|
719 |
+
def inv_key_mapping_layers(key):
|
720 |
+
return re.sub(r"bert.encoder.layers.", "bert.encoder.layer.", key)
|
721 |
+
|
722 |
+
def inv_key_mapping_mlp(key):
|
723 |
+
key = re.sub(
|
724 |
+
r"bert.encoder.layer.(\d+).mlp.fc1.(weight|bias)",
|
725 |
+
r"bert.encoder.layer.\1.intermediate.dense.\2",
|
726 |
+
key,
|
727 |
+
)
|
728 |
+
key = re.sub(
|
729 |
+
r"bert.encoder.layer.(\d+).mlp.fc2.(weight|bias)",
|
730 |
+
r"bert.encoder.layer.\1.output.dense.\2",
|
731 |
+
key,
|
732 |
+
)
|
733 |
+
return key
|
734 |
+
|
735 |
+
def inv_key_mapping_attn(key):
|
736 |
+
return re.sub(
|
737 |
+
r"bert.encoder.layer.(\d+).mixer.out_proj.(weight|bias)",
|
738 |
+
r"bert.encoder.layer.\1.attention.output.dense.\2",
|
739 |
+
key,
|
740 |
+
)
|
741 |
+
|
742 |
+
def inv_key_mapping_decoder_bias(key):
|
743 |
+
return re.sub(r"cls.predictions.decoder.bias", "cls.predictions.bias", key)
|
744 |
+
|
745 |
+
state_dict = OrderedDict((inv_key_mapping_ln(key), value) for key, value in state_dict.items())
|
746 |
+
state_dict = OrderedDict(
|
747 |
+
(inv_key_mapping_ln_gamma_beta(key), value) for key, value in state_dict.items()
|
748 |
+
)
|
749 |
+
state_dict = OrderedDict(
|
750 |
+
(inv_key_mapping_layers(key), value) for key, value in state_dict.items()
|
751 |
+
)
|
752 |
+
state_dict = OrderedDict((inv_key_mapping_mlp(key), value) for key, value in state_dict.items())
|
753 |
+
state_dict = OrderedDict(
|
754 |
+
(inv_key_mapping_attn(key), value) for key, value in state_dict.items()
|
755 |
+
)
|
756 |
+
state_dict = OrderedDict(
|
757 |
+
(inv_key_mapping_decoder_bias(key), value) for key, value in state_dict.items()
|
758 |
+
)
|
759 |
+
|
760 |
+
return state_dict
|