Upload NomicBertForPreTraining
Browse files- config.json +54 -0
- configuration_nomic_bert.py +51 -0
- modeling_hf_nomic_bert.py +881 -0
- pytorch_model.bin +3 -0
config.json
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"activation_function": "swiglu",
|
3 |
+
"architectures": [
|
4 |
+
"NomicBertForPreTraining"
|
5 |
+
],
|
6 |
+
"attn_pdrop": 0.0,
|
7 |
+
"auto_map": {
|
8 |
+
"AutoConfig": "configuration_nomic_bert.NomicBertConfig",
|
9 |
+
"AutoModelForMaskedLM": "modeling_hf_nomic_bert.NomicBertForPreTraining"
|
10 |
+
},
|
11 |
+
"bos_token_id": null,
|
12 |
+
"causal": false,
|
13 |
+
"dense_seq_output": true,
|
14 |
+
"embd_pdrop": 0.1,
|
15 |
+
"eos_token_id": null,
|
16 |
+
"fused_bias_fc": true,
|
17 |
+
"fused_dropout_add_ln": true,
|
18 |
+
"initializer_range": 0.02,
|
19 |
+
"layer_norm_epsilon": 1e-12,
|
20 |
+
"mlp_fc1_bias": false,
|
21 |
+
"mlp_fc2_bias": false,
|
22 |
+
"model_type": "nomic_bert",
|
23 |
+
"n_embd": 768,
|
24 |
+
"n_head": 12,
|
25 |
+
"n_inner": 3072,
|
26 |
+
"n_layer": 12,
|
27 |
+
"n_positions": 0,
|
28 |
+
"pad_vocab_size_multiple": 64,
|
29 |
+
"parallel_block": false,
|
30 |
+
"parallel_block_tied_norm": false,
|
31 |
+
"prenorm": false,
|
32 |
+
"qkv_proj_bias": false,
|
33 |
+
"reorder_and_upcast_attn": false,
|
34 |
+
"resid_pdrop": 0.1,
|
35 |
+
"rotary_emb_base": 1000,
|
36 |
+
"rotary_emb_fraction": 1.0,
|
37 |
+
"rotary_emb_interleaved": false,
|
38 |
+
"rotary_emb_scale_base": null,
|
39 |
+
"scale_attn_by_inverse_layer_idx": false,
|
40 |
+
"scale_attn_weights": true,
|
41 |
+
"summary_activation": null,
|
42 |
+
"summary_first_dropout": 0.1,
|
43 |
+
"summary_proj_to_labels": true,
|
44 |
+
"summary_type": "cls_index",
|
45 |
+
"summary_use_proj": true,
|
46 |
+
"torch_dtype": "float32",
|
47 |
+
"transformers_version": "4.34.0",
|
48 |
+
"type_vocab_size": 2,
|
49 |
+
"use_cache": true,
|
50 |
+
"use_flash_attn": true,
|
51 |
+
"use_rms_norm": false,
|
52 |
+
"use_xentropy": true,
|
53 |
+
"vocab_size": 30528
|
54 |
+
}
|
configuration_nomic_bert.py
ADDED
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import GPT2Config
|
2 |
+
|
3 |
+
|
4 |
+
class NomicBertConfig(GPT2Config):
|
5 |
+
model_type = "nomic_bert"
|
6 |
+
|
7 |
+
def __init__(self,
|
8 |
+
prenorm=False,
|
9 |
+
parallel_block=False,
|
10 |
+
parallel_block_tied_norm=False,
|
11 |
+
rotary_emb_fraction=0.0,
|
12 |
+
fused_dropout_add_ln=False,
|
13 |
+
fused_bias_fc=False,
|
14 |
+
use_flash_attn=False,
|
15 |
+
use_xentropy=False,
|
16 |
+
qkv_proj_bias=True,
|
17 |
+
rotary_emb_base=1000,
|
18 |
+
rotary_emb_scale_base=None,
|
19 |
+
rotary_emb_interleaved=False,
|
20 |
+
mlp_fc1_bias=True,
|
21 |
+
mlp_fc2_bias=True,
|
22 |
+
use_rms_norm=False,
|
23 |
+
causal=False,
|
24 |
+
type_vocab_size=2,
|
25 |
+
dense_seq_output=True,
|
26 |
+
pad_vocab_size_multiple=1,
|
27 |
+
tie_word_embeddings=True,
|
28 |
+
**kwargs,
|
29 |
+
):
|
30 |
+
self.prenorm = prenorm
|
31 |
+
self.parallel_block = parallel_block
|
32 |
+
self.parallel_block_tied_norm = parallel_block_tied_norm
|
33 |
+
self.rotary_emb_fraction = rotary_emb_fraction
|
34 |
+
self.tie_word_embeddings = tie_word_embeddings
|
35 |
+
self.fused_dropout_add_ln = fused_dropout_add_ln
|
36 |
+
self.fused_bias_fc = fused_bias_fc
|
37 |
+
self.use_flash_attn = use_flash_attn
|
38 |
+
self.use_xentropy = use_xentropy
|
39 |
+
self.qkv_proj_bias = qkv_proj_bias
|
40 |
+
self.rotary_emb_base = rotary_emb_base
|
41 |
+
self.rotary_emb_scale_base = rotary_emb_scale_base
|
42 |
+
self.rotary_emb_interleaved = rotary_emb_interleaved
|
43 |
+
self.mlp_fc1_bias = mlp_fc1_bias
|
44 |
+
self.mlp_fc2_bias = mlp_fc2_bias
|
45 |
+
self.use_rms_norm = use_rms_norm
|
46 |
+
self.causal = causal
|
47 |
+
self.type_vocab_size = type_vocab_size
|
48 |
+
self.dense_seq_output = dense_seq_output
|
49 |
+
self.pad_vocab_size_multiple = pad_vocab_size_multiple
|
50 |
+
|
51 |
+
super().__init__(**kwargs)
|
modeling_hf_nomic_bert.py
ADDED
@@ -0,0 +1,881 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 os
|
9 |
+
import logging
|
10 |
+
from functools import partial
|
11 |
+
from typing import Optional, List, Tuple, Union
|
12 |
+
|
13 |
+
import torch
|
14 |
+
import torch.nn as nn
|
15 |
+
import torch.nn.functional as F
|
16 |
+
from einops import rearrange, repeat
|
17 |
+
from transformers import GPT2Config, PreTrainedModel
|
18 |
+
from transformers.models.bert.modeling_bert import (
|
19 |
+
BaseModelOutputWithPoolingAndCrossAttentions,
|
20 |
+
BertForPreTrainingOutput,
|
21 |
+
SequenceClassifierOutput
|
22 |
+
)
|
23 |
+
|
24 |
+
from contrastors.models.encoder.configuration_nomic_bert import NomicBertConfig
|
25 |
+
from contrastors.models.model_utils import state_dict_from_pretrained, filter_shapes
|
26 |
+
from contrastors.models.encoder.bert import remap_bert_state_dict
|
27 |
+
|
28 |
+
logger = logging.getLogger(__name__)
|
29 |
+
|
30 |
+
|
31 |
+
class NomicBertPreTrainedModel(PreTrainedModel):
|
32 |
+
"""An abstract class to handle weights initialization and
|
33 |
+
a simple interface for dowloading and loading pretrained models.
|
34 |
+
"""
|
35 |
+
config_class = NomicBertConfig
|
36 |
+
base_model_prefix = "model"
|
37 |
+
supports_gradient_checkpointing = True
|
38 |
+
_no_split_modules = ["Block"]
|
39 |
+
_skip_keys_device_placement = "past_key_values"
|
40 |
+
|
41 |
+
def __init__(self, config, *inputs, **kwargs):
|
42 |
+
super().__init__(config)
|
43 |
+
if not isinstance(config, GPT2Config):
|
44 |
+
raise ValueError(
|
45 |
+
"Parameter config in `{}(config)` should be an instance of class `GPT2Config`. "
|
46 |
+
"To create a model from a Google pretrained model use "
|
47 |
+
"`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
|
48 |
+
self.__class__.__name__, self.__class__.__name__
|
49 |
+
)
|
50 |
+
)
|
51 |
+
self.config = config
|
52 |
+
|
53 |
+
@classmethod
|
54 |
+
def from_pretrained(cls, model_name, config=None, *inputs, **kwargs):
|
55 |
+
"""
|
56 |
+
Instantiate a NomicBertPreTrainedModel from a pre-trained model file or a pytorch state dict.
|
57 |
+
Download and cache the pre-trained model file if needed.
|
58 |
+
|
59 |
+
Params:
|
60 |
+
pretrained_model_name_or_path: either:
|
61 |
+
- a path or url to a pretrained model archive containing:
|
62 |
+
. `bert_config.json` a configuration file for the model
|
63 |
+
. `pytorch_model.bin` a PyTorch dump of a NomicBertForPretraining instance
|
64 |
+
- a path or url to a pretrained model archive containing:
|
65 |
+
. `bert_config.json` a configuration file for the model
|
66 |
+
. `model.chkpt` a TensorFlow checkpoint
|
67 |
+
*inputs, **kwargs: additional input for the specific NomicBert class
|
68 |
+
(ex: num_labels for NomicBertForSequenceClassification)
|
69 |
+
"""
|
70 |
+
# Instantiate model.
|
71 |
+
if config is None:
|
72 |
+
config = cls.config_class.from_pretrained(model_name)
|
73 |
+
remove_cls = cls != NomicBertForPreTraining
|
74 |
+
remove_bert_prefix = cls != NomicBertForPreTraining
|
75 |
+
ignore_mismatched_shapes = kwargs.pop("ignore_mismatched_sizes", False)
|
76 |
+
model = cls(config, *inputs, **kwargs)
|
77 |
+
# TODO: fix this
|
78 |
+
# Assuming we know what we're doing when loading from disk
|
79 |
+
# Prob a bad assumption but i'm tired and want to train this asap
|
80 |
+
if os.path.exists(model_name):
|
81 |
+
state_dict = torch.load(f"{model_name}/pytorch_model.bin")
|
82 |
+
if ignore_mismatched_shapes:
|
83 |
+
state_dict = filter_shapes(state_dict, model)
|
84 |
+
load_return = model.load_state_dict(state_dict, strict=False)
|
85 |
+
else:
|
86 |
+
# TODO: can probably check config class and see if we need to remap from a bert model
|
87 |
+
state_dict = state_dict_from_pretrained(model_name)
|
88 |
+
state_dict = remap_bert_state_dict(state_dict,
|
89 |
+
config,
|
90 |
+
remove_bert=remove_bert_prefix,
|
91 |
+
remove_cls_weights=remove_cls,
|
92 |
+
add_pooling_layer=getattr(config, "add_pooling_layer", False)
|
93 |
+
)
|
94 |
+
if ignore_mismatched_shapes:
|
95 |
+
state_dict = filter_shapes(state_dict, model)
|
96 |
+
|
97 |
+
load_return = model.load_state_dict(
|
98 |
+
state_dict,
|
99 |
+
strict=True
|
100 |
+
)
|
101 |
+
logger.info(load_return)
|
102 |
+
return model
|
103 |
+
|
104 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
105 |
+
if isinstance(module, NomicBertEncoder):
|
106 |
+
module.gradient_checkpointing = value
|
107 |
+
|
108 |
+
|
109 |
+
# https://github.com/huggingface/transformers/blob/7032e0203262ebb2ebf55da8d2e01f873973e835/src/transformers/models/bert/modeling_bert.py#L748
|
110 |
+
def _init_weights(module, initializer_range=0.02):
|
111 |
+
if isinstance(module, nn.Linear):
|
112 |
+
nn.init.normal_(module.weight, std=initializer_range)
|
113 |
+
if module.bias is not None:
|
114 |
+
nn.init.zeros_(module.bias)
|
115 |
+
elif isinstance(module, nn.Embedding):
|
116 |
+
nn.init.normal_(module.weight, std=initializer_range)
|
117 |
+
if module.padding_idx is not None:
|
118 |
+
nn.init.zeros_(module.weight[module.padding_idx])
|
119 |
+
|
120 |
+
|
121 |
+
class NomicBertEmbeddings(nn.Module):
|
122 |
+
def __init__(
|
123 |
+
self,
|
124 |
+
config
|
125 |
+
):
|
126 |
+
"""
|
127 |
+
If max_position_embeddings <= 0, there's no position embeddings
|
128 |
+
If type_vocab_size <= 0, there's no token type embeddings
|
129 |
+
"""
|
130 |
+
super().__init__()
|
131 |
+
self.word_embeddings = nn.Embedding(
|
132 |
+
config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
|
133 |
+
)
|
134 |
+
self.max_position_embeddings = config.max_position_embeddings
|
135 |
+
self.type_vocab_size = config.type_vocab_size
|
136 |
+
if self.max_position_embeddings > 0:
|
137 |
+
self.position_embeddings = nn.Embedding(
|
138 |
+
config.max_position_embeddings, config.hidden_size,
|
139 |
+
)
|
140 |
+
if self.type_vocab_size > 0:
|
141 |
+
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
|
142 |
+
|
143 |
+
def forward(self, input_ids, position_ids=None, token_type_ids=None):
|
144 |
+
"""
|
145 |
+
input_ids: (batch, seqlen)
|
146 |
+
position_ids: (batch, seqlen)
|
147 |
+
token_type_ids: (batch, seqlen)
|
148 |
+
"""
|
149 |
+
batch_size, seqlen = input_ids.shape
|
150 |
+
embeddings = self.word_embeddings(input_ids)
|
151 |
+
|
152 |
+
if self.type_vocab_size > 0:
|
153 |
+
if token_type_ids is None:
|
154 |
+
token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
|
155 |
+
token_type_embeddings = self.token_type_embeddings(token_type_ids)
|
156 |
+
embeddings = embeddings + token_type_embeddings
|
157 |
+
|
158 |
+
if self.max_position_embeddings > 0:
|
159 |
+
if position_ids is None:
|
160 |
+
position_ids = torch.arange(seqlen, dtype=torch.long, device=input_ids.device)
|
161 |
+
position_embeddings = self.position_embeddings(position_ids)
|
162 |
+
embeddings = embeddings + position_embeddings
|
163 |
+
return embeddings
|
164 |
+
|
165 |
+
class NomicBertMLP(nn.Module):
|
166 |
+
def __init__(
|
167 |
+
self,
|
168 |
+
in_features,
|
169 |
+
hidden_features=None,
|
170 |
+
out_features=None,
|
171 |
+
activation=F.gelu,
|
172 |
+
bias1=True,
|
173 |
+
bias2=True,
|
174 |
+
return_residual=False,
|
175 |
+
fused_bias_fc=False,
|
176 |
+
):
|
177 |
+
super().__init__()
|
178 |
+
out_features = out_features if out_features is not None else in_features
|
179 |
+
hidden_features = hidden_features if hidden_features is not None else in_features * 4
|
180 |
+
self.return_residual = return_residual
|
181 |
+
self.fc1 = nn.Linear(in_features, hidden_features, bias=bias1)
|
182 |
+
approximate = (
|
183 |
+
"tanh"
|
184 |
+
if activation in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
|
185 |
+
else "none"
|
186 |
+
)
|
187 |
+
self.activation = nn.GELU(approximate=approximate) if activation == "gelu" else activation
|
188 |
+
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2)
|
189 |
+
|
190 |
+
def forward(self, x):
|
191 |
+
y = self.fc1(x)
|
192 |
+
y = self.activation(y)
|
193 |
+
y = self.fc2(y)
|
194 |
+
return y if not self.return_residual else (y, x)
|
195 |
+
|
196 |
+
|
197 |
+
class NomciBertGatedMLP(nn.Module):
|
198 |
+
def __init__(
|
199 |
+
self,
|
200 |
+
in_features,
|
201 |
+
hidden_features=None,
|
202 |
+
out_features=None,
|
203 |
+
activation=F.sigmoid,
|
204 |
+
bias1=True,
|
205 |
+
bias2=True,
|
206 |
+
multiple_of=256,
|
207 |
+
return_residual=False,
|
208 |
+
fused_bias_fc=True,
|
209 |
+
device=None,
|
210 |
+
dtype=None,
|
211 |
+
):
|
212 |
+
super().__init__()
|
213 |
+
out_features = out_features if out_features is not None else in_features
|
214 |
+
hidden_features = (
|
215 |
+
hidden_features if hidden_features is not None else int(8 * in_features / 3)
|
216 |
+
)
|
217 |
+
hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
|
218 |
+
self.return_residual = return_residual
|
219 |
+
|
220 |
+
self.fc11 = nn.Linear(in_features, hidden_features, bias=bias1)
|
221 |
+
self.fc12 = nn.Linear(in_features, hidden_features, bias=bias1)
|
222 |
+
self.activation = activation
|
223 |
+
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2)
|
224 |
+
|
225 |
+
def forward(self, x):
|
226 |
+
y = self.fc11(x)
|
227 |
+
gate = self.fc12(x)
|
228 |
+
if self.activation == F.sigmoid: # Special case for GLU
|
229 |
+
y = F.glu(torch.cat([y, gate], dim=-1), dim=-1)
|
230 |
+
else:
|
231 |
+
y = y * self.activation(gate)
|
232 |
+
y = self.fc2(y)
|
233 |
+
return y if not self.return_residual else (y, x)
|
234 |
+
|
235 |
+
|
236 |
+
def rotate_half(x, interleaved=False):
|
237 |
+
if not interleaved:
|
238 |
+
x1, x2 = x.chunk(2, dim=-1)
|
239 |
+
return torch.cat((-x2, x1), dim=-1)
|
240 |
+
else:
|
241 |
+
x1, x2 = x[..., ::2], x[..., 1::2]
|
242 |
+
return rearrange(torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2)
|
243 |
+
|
244 |
+
|
245 |
+
def apply_rotary_emb(x, cos, sin, offset=0, interleaved=False):
|
246 |
+
"""
|
247 |
+
x: (batch_size, seqlen, nheads, headdim)
|
248 |
+
cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
|
249 |
+
"""
|
250 |
+
ro_dim = cos.shape[-1] * 2
|
251 |
+
assert ro_dim <= x.shape[-1]
|
252 |
+
cos, sin = (
|
253 |
+
cos[offset: offset + x.shape[1]],
|
254 |
+
sin[offset: offset + x.shape[1]],
|
255 |
+
)
|
256 |
+
cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
257 |
+
sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
258 |
+
return torch.cat(
|
259 |
+
[x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]],
|
260 |
+
dim=-1,
|
261 |
+
)
|
262 |
+
|
263 |
+
|
264 |
+
class NomicBertRotaryEmbedding(nn.Module):
|
265 |
+
def __init__(
|
266 |
+
self,
|
267 |
+
dim: int,
|
268 |
+
base=10000.0,
|
269 |
+
interleaved=False,
|
270 |
+
scale_base=None,
|
271 |
+
pos_idx_in_fp32=True,
|
272 |
+
device=None,
|
273 |
+
):
|
274 |
+
"""
|
275 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
|
276 |
+
of 1st half and 2nd half (GPT-NeoX style).
|
277 |
+
pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
|
278 |
+
otherwise they might be in lower precision.
|
279 |
+
This option was added because previously (before 2023-07-02), when we construct
|
280 |
+
the position indices, we use the dtype of self.inv_freq. In most cases this would
|
281 |
+
be fp32, but if the model is trained in pure bf16 (not mixed precision), then
|
282 |
+
self.inv_freq would be bf16, and the position indices are also in bf16.
|
283 |
+
Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
|
284 |
+
embeddings for some positions will coincide.
|
285 |
+
To maintain compatibility with models previously trained in pure bf16,
|
286 |
+
we add this option.
|
287 |
+
"""
|
288 |
+
super().__init__()
|
289 |
+
self.dim = dim
|
290 |
+
self.base = float(base)
|
291 |
+
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
292 |
+
# Generate and save the inverse frequency buffer (non trainable)
|
293 |
+
inv_freq = self._compute_inv_freq(device)
|
294 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
295 |
+
self.interleaved = interleaved
|
296 |
+
self.scale_base = scale_base
|
297 |
+
|
298 |
+
self._seq_len_cached = 0
|
299 |
+
self._cos_cached = None
|
300 |
+
self._sin_cached = None
|
301 |
+
self._cos_k_cached = None
|
302 |
+
self._sin_k_cached = None
|
303 |
+
|
304 |
+
def _compute_inv_freq(self, device=None):
|
305 |
+
return 1.0 / (
|
306 |
+
self.base
|
307 |
+
** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim)
|
308 |
+
)
|
309 |
+
|
310 |
+
def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
|
311 |
+
# Reset the tables if the sequence length has changed,
|
312 |
+
# if we're on a new device (possibly due to tracing for instance),
|
313 |
+
# or if we're switching from inference mode to training
|
314 |
+
if (
|
315 |
+
seqlen > self._seq_len_cached
|
316 |
+
or self._cos_cached is None
|
317 |
+
or self._cos_cached.device != device
|
318 |
+
or self._cos_cached.dtype != dtype
|
319 |
+
or (self.training and self._cos_cached.is_inference())
|
320 |
+
):
|
321 |
+
self._seq_len_cached = seqlen
|
322 |
+
# We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
|
323 |
+
# And the output of arange can be quite large, so bf16 would lose a lot of precision.
|
324 |
+
# However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
|
325 |
+
if self.pos_idx_in_fp32:
|
326 |
+
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
327 |
+
# We want fp32 here as well since inv_freq will be multiplied with t, and the output
|
328 |
+
# will be large. Having it in bf16 will lose a lot of precision and cause the
|
329 |
+
# cos & sin output to change significantly.
|
330 |
+
# We want to recompute self.inv_freq if it was not loaded in fp32
|
331 |
+
if self.inv_freq.dtype != torch.float32:
|
332 |
+
inv_freq = self._compute_inv_freq(device=device)
|
333 |
+
else:
|
334 |
+
inv_freq = self.inv_freq
|
335 |
+
else:
|
336 |
+
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
337 |
+
inv_freq = self.inv_freq
|
338 |
+
# Don't do einsum, it converts fp32 to fp16 under AMP
|
339 |
+
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
340 |
+
freqs = torch.outer(t, inv_freq)
|
341 |
+
self._cos_cached = torch.cos(freqs).to(dtype)
|
342 |
+
self._sin_cached = torch.sin(freqs).to(dtype)
|
343 |
+
|
344 |
+
def forward(
|
345 |
+
self,
|
346 |
+
qkv: torch.Tensor,
|
347 |
+
kv: Optional[torch.Tensor] = None,
|
348 |
+
seqlen_offset: Union[int, torch.Tensor] = 0,
|
349 |
+
max_seqlen: Optional[int] = None,
|
350 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
351 |
+
"""
|
352 |
+
qkv: (batch, seqlen, 3, nheads, headdim) if kv is none,
|
353 |
+
else it's just q of shape (batch, seqlen, nheads, headdim)
|
354 |
+
kv: (batch, seqlen, 2, nheads, headdim)
|
355 |
+
seqlen_offset: (batch_size,) or int. Each sequence in x is shifted by this amount.
|
356 |
+
Most commonly used in inference when we have KV cache.
|
357 |
+
If it's a tensor of shape (batch_size,), then to update the cos / sin cache, one
|
358 |
+
should pass in max_seqlen, which will update the cos / sin cache up to that length.
|
359 |
+
Apply rotary embedding *inplace* to qkv and / or kv.
|
360 |
+
"""
|
361 |
+
seqlen = qkv.shape[1]
|
362 |
+
if max_seqlen is not None:
|
363 |
+
self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
|
364 |
+
elif isinstance(seqlen_offset, int):
|
365 |
+
self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
|
366 |
+
|
367 |
+
q_rot = apply_rotary_emb(qkv[:, :, 0], self._cos_cached, self._sin_cached, seqlen_offset, self.interleaved)
|
368 |
+
k_rot = apply_rotary_emb(qkv[:, :, 1], self._cos_cached, self._sin_cached, seqlen_offset, self.interleaved)
|
369 |
+
return torch.stack((q_rot, k_rot, qkv[:, :, 2]), dim=2)
|
370 |
+
|
371 |
+
|
372 |
+
|
373 |
+
class NomicBertAttention(nn.Module):
|
374 |
+
"""Multi-head self-attention and cross-attention"""
|
375 |
+
|
376 |
+
def __init__(
|
377 |
+
self,
|
378 |
+
config,
|
379 |
+
) -> None:
|
380 |
+
"""
|
381 |
+
num_heads_kv: can be used to toggle MQA / GQA. If None, use num_heads.
|
382 |
+
return_residual: whether to return the input x along with the output. This is for
|
383 |
+
performance reason: for post-norm architecture, returning the input allows us
|
384 |
+
to fuse the backward of nn.Linear with the residual connection.
|
385 |
+
"""
|
386 |
+
super().__init__()
|
387 |
+
self.embed_dim = config.n_embd
|
388 |
+
self.use_flash_attn = config.use_flash_attn
|
389 |
+
self.fused_bias_fc = config.fused_bias_fc
|
390 |
+
|
391 |
+
self.num_heads = config.n_head
|
392 |
+
self.num_heads_kv = config.num_heads_kv if getattr(config, "num_heads_kv", None) is not None else self.num_heads
|
393 |
+
assert self.embed_dim % self.num_heads == 0, "embed_dim must be divisible by num_heads"
|
394 |
+
self.head_dim = self.embed_dim // self.num_heads
|
395 |
+
# we don't really support mqa / gqa for now
|
396 |
+
qkv_dim = self.head_dim * (self.num_heads + 2 * self.num_heads_kv)
|
397 |
+
|
398 |
+
self.register_buffer(
|
399 |
+
"norm_factor",
|
400 |
+
torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()),
|
401 |
+
persistent=False,
|
402 |
+
)
|
403 |
+
|
404 |
+
self.rotary_emb_dim = self.head_dim * config.rotary_emb_fraction
|
405 |
+
if self.rotary_emb_dim > 0:
|
406 |
+
self.rotary_emb = NomicBertRotaryEmbedding(
|
407 |
+
self.rotary_emb_dim,
|
408 |
+
base=config.rotary_emb_base,
|
409 |
+
scale_base=config.rotary_emb_scale_base,
|
410 |
+
interleaved=config.rotary_emb_interleaved,
|
411 |
+
)
|
412 |
+
# bug in xformers: https://github.com/facebookresearch/xformers/issues/841
|
413 |
+
# uses the head dimension instead of the sequence dimension
|
414 |
+
self.rotary_head_dim = getattr(config, "rotary_head_dim", False)
|
415 |
+
|
416 |
+
self.Wqkv = nn.Linear(self.embed_dim, qkv_dim, bias=config.qkv_proj_bias)
|
417 |
+
|
418 |
+
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.qkv_proj_bias)
|
419 |
+
self.causal = config.causal
|
420 |
+
self.drop = nn.Dropout(config.attn_pdrop)
|
421 |
+
|
422 |
+
def forward(
|
423 |
+
self,
|
424 |
+
hidden_states: torch.Tensor,
|
425 |
+
attention_mask: Optional[torch.Tensor] = None,
|
426 |
+
position_ids: Optional[torch.LongTensor] = None,
|
427 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
428 |
+
output_attentions: bool = False,
|
429 |
+
use_cache: bool = False,
|
430 |
+
is_padded_inputs: Optional[bool] = True,
|
431 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
432 |
+
max_seq_len: Optional[int] = None,
|
433 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
434 |
+
|
435 |
+
has_layer_past = past_key_value is not None
|
436 |
+
|
437 |
+
if has_layer_past:
|
438 |
+
past_key_value = past_key_value[0]
|
439 |
+
past_len = past_key_value[1]
|
440 |
+
else:
|
441 |
+
past_len = 0
|
442 |
+
|
443 |
+
qkv = self.Wqkv(hidden_states)
|
444 |
+
qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
|
445 |
+
|
446 |
+
past_key_value = (past_key_value, past_len + qkv.size(1)) if use_cache else None
|
447 |
+
|
448 |
+
if self.rotary_emb_dim > 0:
|
449 |
+
if self.rotary_head_dim:
|
450 |
+
qkv = rearrange(qkv, "b s three h d -> b h three s d")
|
451 |
+
qkv = self.rotary_emb(qkv, seqlen_offset=past_len)
|
452 |
+
|
453 |
+
if self.rotary_head_dim:
|
454 |
+
qkv = rearrange(qkv, "b h three s d -> b s three h d")
|
455 |
+
|
456 |
+
query, key, value = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2]
|
457 |
+
|
458 |
+
query = query.permute(0, 2, 1, 3)
|
459 |
+
key = key.permute(0, 2, 1, 3)
|
460 |
+
value = value.permute(0, 2, 1, 3)
|
461 |
+
|
462 |
+
attention_scores = torch.matmul(query, key.transpose(-1, -2)) / self.norm_factor
|
463 |
+
if attention_mask is not None:
|
464 |
+
attention_scores = attention_scores + attention_mask
|
465 |
+
|
466 |
+
attentions_probs = F.softmax(attention_scores, dim=-1)
|
467 |
+
attentions_probs = self.drop(attentions_probs)
|
468 |
+
|
469 |
+
attn_output = torch.matmul(attentions_probs, value)
|
470 |
+
attn_output = rearrange(attn_output.permute(0, 2, 1, 3), "... h d -> ... (h d)")
|
471 |
+
|
472 |
+
attn_output = self.out_proj(attn_output)
|
473 |
+
|
474 |
+
return attn_output
|
475 |
+
|
476 |
+
|
477 |
+
class NomicBertBlock(nn.Module):
|
478 |
+
def __init__(
|
479 |
+
self,
|
480 |
+
config,
|
481 |
+
):
|
482 |
+
super().__init__()
|
483 |
+
self.prenorm = config.prenorm
|
484 |
+
self.fused_dropout_add_ln = config.fused_dropout_add_ln
|
485 |
+
|
486 |
+
self.attn = NomicBertAttention(config)
|
487 |
+
activation = (
|
488 |
+
F.sigmoid
|
489 |
+
if config.activation_function == "glu"
|
490 |
+
else (F.silu if config.activation_function == "swiglu" else F.gelu)
|
491 |
+
)
|
492 |
+
if config.activation_function in ["glu", "swiglu", "geglu"]:
|
493 |
+
self.mlp = NomciBertGatedMLP(config.n_embd, hidden_features=config.n_inner, bias1=config.mlp_fc1_bias, bias2=config.mlp_fc2_bias, activation=activation, fused_bias_fc=config.fused_bias_fc)
|
494 |
+
else:
|
495 |
+
self.mlp = NomicBertMLP(config.n_embd, hidden_features=config.n_inner, bias1=config.mlp_fc1_bias, bias2=config.mlp_fc2_bias, activation=activation, fused_bias_fc=config.fused_bias_fc)
|
496 |
+
|
497 |
+
self.dropout1 = nn.Dropout(config.resid_pdrop)
|
498 |
+
self.norm1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
499 |
+
self.norm2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
500 |
+
self.dropout2 = nn.Dropout(config.resid_pdrop)
|
501 |
+
|
502 |
+
def forward(
|
503 |
+
self,
|
504 |
+
hidden_states: torch.Tensor,
|
505 |
+
hidden_states2: torch.Tensor,
|
506 |
+
residual: Optional[torch.Tensor] = None,
|
507 |
+
attention_mask: Optional[torch.Tensor] = None,
|
508 |
+
position_ids: Optional[torch.LongTensor] = None,
|
509 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
510 |
+
is_padded_inputs: Optional[bool] = True,
|
511 |
+
output_attentions: Optional[bool] = False,
|
512 |
+
use_cache: Optional[bool] = False,
|
513 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
514 |
+
max_seq_len: Optional[int] = None,
|
515 |
+
):
|
516 |
+
r"""Pass the input through the encoder layer.
|
517 |
+
|
518 |
+
Args:
|
519 |
+
hidden_states: the sequence to the encoder layer (required).
|
520 |
+
residual: if postnorm, residual=None, If prenorm, hidden_states = Attn/MLP(LN(residual))
|
521 |
+
mixer_subset: for cross-attention only. If not None, will take a subset of x
|
522 |
+
before applying the query projection. Useful for e.g., ViT where we only care
|
523 |
+
about the CLS token in the last layer.
|
524 |
+
"""
|
525 |
+
if self.prenorm:
|
526 |
+
dropped = self.dropout1(hidden_states)
|
527 |
+
residual = (dropped + residual) if residual is not None else dropped
|
528 |
+
hidden_states = self.norm1(residual.to(dtype=self.norm1.weight.dtype))
|
529 |
+
hidden_states = self.attn(hidden_states, attention_mask=attention_mask, is_padded_inputs=is_padded_inputs, cu_seqlens=cu_seqlens, max_seq_len=max_seq_len)
|
530 |
+
|
531 |
+
dropped = self.dropout2(hidden_states)
|
532 |
+
residual = (dropped + residual) if residual is not None else dropped
|
533 |
+
hidden_states = self.norm2(residual.to(dtype=self.norm2.weight.dtype))
|
534 |
+
hidden_states = self.mlp(hidden_states)
|
535 |
+
|
536 |
+
return hidden_states, None, residual
|
537 |
+
else:
|
538 |
+
assert residual is None
|
539 |
+
attn_outputs = self.attn(hidden_states,
|
540 |
+
attention_mask=attention_mask,
|
541 |
+
is_padded_inputs=is_padded_inputs,
|
542 |
+
cu_seqlens=cu_seqlens,
|
543 |
+
max_seq_len=max_seq_len)
|
544 |
+
hidden_states = self.norm1(
|
545 |
+
(self.dropout1(attn_outputs) + hidden_states).to(
|
546 |
+
dtype=self.norm1.weight.dtype
|
547 |
+
)
|
548 |
+
)
|
549 |
+
mlp_out = self.mlp(hidden_states)
|
550 |
+
|
551 |
+
hidden_states = self.norm2(
|
552 |
+
(self.dropout2(mlp_out) + hidden_states).to(
|
553 |
+
dtype=self.norm2.weight.dtype
|
554 |
+
)
|
555 |
+
)
|
556 |
+
return hidden_states, None, None
|
557 |
+
|
558 |
+
|
559 |
+
class NomicBertEncoder(nn.Module):
|
560 |
+
def __init__(self, config: GPT2Config):
|
561 |
+
super().__init__()
|
562 |
+
self.layers = nn.ModuleList(
|
563 |
+
[NomicBertBlock(config) for _ in range(config.n_layer)]
|
564 |
+
)
|
565 |
+
self.gradient_checkpointing = False
|
566 |
+
self.config = config
|
567 |
+
|
568 |
+
def forward(self,
|
569 |
+
hidden_states: torch.LongTensor = None,
|
570 |
+
attention_mask: Optional[torch.Tensor] = None,
|
571 |
+
position_ids: Optional[torch.LongTensor] = None,
|
572 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
573 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
574 |
+
use_cache: Optional[bool] = None,
|
575 |
+
output_attentions: Optional[bool] = None,
|
576 |
+
output_hidden_states: Optional[bool] = None,
|
577 |
+
return_dict: Optional[bool] = None,
|
578 |
+
is_padded_inputs: Optional[bool] = True,):
|
579 |
+
|
580 |
+
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
581 |
+
This means that we only compute the last layer output for these tokens.
|
582 |
+
subset_mask: (batch, seqlen), dtype=torch.bool
|
583 |
+
"""
|
584 |
+
hidden_states2 = None
|
585 |
+
residual = None
|
586 |
+
|
587 |
+
|
588 |
+
for _, layer in enumerate(self.layers):
|
589 |
+
if self.gradient_checkpointing and self.training:
|
590 |
+
|
591 |
+
def create_custom_forward(module):
|
592 |
+
def custom_forward(*inputs):
|
593 |
+
# None for past_key_value
|
594 |
+
return module(*inputs)
|
595 |
+
|
596 |
+
return custom_forward
|
597 |
+
|
598 |
+
hidden_states, hidden_states2, residual = torch.utils.checkpoint.checkpoint(
|
599 |
+
create_custom_forward(layer),
|
600 |
+
hidden_states,
|
601 |
+
hidden_states2,
|
602 |
+
residual,
|
603 |
+
attention_mask,
|
604 |
+
None,
|
605 |
+
None,
|
606 |
+
is_padded_inputs,
|
607 |
+
# if you freeze ANY layers, you need `use_reentrant=False`
|
608 |
+
# https://github.com/huggingface/transformers/issues/21381
|
609 |
+
# https://discuss.pytorch.org/t/checkpoint-with-no-grad-requiring-inputs-problem/19117/7
|
610 |
+
use_reentrant=False,
|
611 |
+
)
|
612 |
+
|
613 |
+
else:
|
614 |
+
hidden_states, hidden_states2, residual = layer(
|
615 |
+
hidden_states,
|
616 |
+
hidden_states2,
|
617 |
+
residual,
|
618 |
+
attention_mask,
|
619 |
+
position_ids,
|
620 |
+
None,
|
621 |
+
is_padded_inputs,
|
622 |
+
output_attentions,
|
623 |
+
use_cache,
|
624 |
+
)
|
625 |
+
return hidden_states
|
626 |
+
|
627 |
+
|
628 |
+
class NomicBertPooler(nn.Module):
|
629 |
+
def __init__(self, config):
|
630 |
+
super().__init__()
|
631 |
+
self.dense = nn.Linear(config.n_embd, config.n_embd)
|
632 |
+
self.activation = nn.Tanh()
|
633 |
+
|
634 |
+
def forward(self, hidden_states, pool=True):
|
635 |
+
# We "pool" the model by simply taking the hidden state corresponding
|
636 |
+
# to the first token.
|
637 |
+
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
638 |
+
pooled_output = self.dense(first_token_tensor)
|
639 |
+
pooled_output = self.activation(pooled_output)
|
640 |
+
return pooled_output
|
641 |
+
|
642 |
+
|
643 |
+
class NomicBertPredictionHeadTransform(nn.Module):
|
644 |
+
def __init__(self, config):
|
645 |
+
super().__init__()
|
646 |
+
self.dense = nn.Linear(config.n_embd, config.n_embd, bias=config.mlp_fc1_bias)
|
647 |
+
approximate = (
|
648 |
+
"tanh"
|
649 |
+
if config.activation_function in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
|
650 |
+
else "none"
|
651 |
+
)
|
652 |
+
if config.activation_function == "swiglu":
|
653 |
+
self.transform_act_fn = F.silu
|
654 |
+
else:
|
655 |
+
self.transform_act_fn = nn.GELU(approximate=approximate)
|
656 |
+
|
657 |
+
self.layer_norm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
658 |
+
|
659 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
660 |
+
hidden_states = self.dense(hidden_states)
|
661 |
+
hidden_states = self.transform_act_fn(hidden_states)
|
662 |
+
hidden_states = self.layer_norm(hidden_states)
|
663 |
+
|
664 |
+
return hidden_states
|
665 |
+
|
666 |
+
|
667 |
+
class NomicBertLMPredictionHead(nn.Module):
|
668 |
+
def __init__(self, config):
|
669 |
+
super().__init__()
|
670 |
+
|
671 |
+
self.transform = NomicBertPredictionHeadTransform(config)
|
672 |
+
|
673 |
+
self.decoder = nn.Linear(config.n_embd, config.vocab_size, bias=config.mlp_fc1_bias)
|
674 |
+
|
675 |
+
def forward(self, hidden_states):
|
676 |
+
hidden_states = self.transform(hidden_states)
|
677 |
+
hidden_states = self.decoder(hidden_states)
|
678 |
+
return hidden_states
|
679 |
+
|
680 |
+
|
681 |
+
class NomicBertPreTrainingHeads(nn.Module):
|
682 |
+
def __init__(self, config):
|
683 |
+
super().__init__()
|
684 |
+
self.predictions = NomicBertLMPredictionHead(config)
|
685 |
+
|
686 |
+
def forward(self, sequence_output):
|
687 |
+
prediction_scores = self.predictions(sequence_output)
|
688 |
+
return prediction_scores
|
689 |
+
|
690 |
+
|
691 |
+
class NomicBertModel(NomicBertPreTrainedModel):
|
692 |
+
def __init__(self, config: GPT2Config, add_pooling_layer=True):
|
693 |
+
super().__init__(config)
|
694 |
+
self.pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
695 |
+
if config.vocab_size % self.pad_vocab_size_multiple != 0:
|
696 |
+
config.vocab_size += self.pad_vocab_size_multiple - (
|
697 |
+
config.vocab_size % self.pad_vocab_size_multiple
|
698 |
+
)
|
699 |
+
|
700 |
+
assert config.activation_function in ["gelu", "gelu_new", "gelu_fast", "gelu_pytorch_tanh", "swiglu", "geglu", "glu"]
|
701 |
+
|
702 |
+
self.embeddings = NomicBertEmbeddings(
|
703 |
+
config
|
704 |
+
)
|
705 |
+
self.emb_drop = nn.Dropout(config.resid_pdrop)
|
706 |
+
self.emb_ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
707 |
+
self.encoder = NomicBertEncoder(config)
|
708 |
+
self.pooler = NomicBertPooler(config) if add_pooling_layer else None
|
709 |
+
|
710 |
+
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
711 |
+
|
712 |
+
def forward(
|
713 |
+
self,
|
714 |
+
input_ids,
|
715 |
+
position_ids=None,
|
716 |
+
token_type_ids=None,
|
717 |
+
attention_mask=None,
|
718 |
+
):
|
719 |
+
if token_type_ids is None:
|
720 |
+
token_type_ids = torch.zeros_like(input_ids)
|
721 |
+
hidden_states = self.embeddings(
|
722 |
+
input_ids, position_ids=position_ids, token_type_ids=token_type_ids
|
723 |
+
)
|
724 |
+
hidden_states = self.emb_ln(hidden_states)
|
725 |
+
hidden_states = self.emb_drop(hidden_states)
|
726 |
+
|
727 |
+
attention_mask = self.get_extended_attention_mask(attention_mask, input_ids.shape)
|
728 |
+
sequence_output = self.encoder(
|
729 |
+
hidden_states, attention_mask=attention_mask
|
730 |
+
)
|
731 |
+
|
732 |
+
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
|
733 |
+
|
734 |
+
return BaseModelOutputWithPoolingAndCrossAttentions(
|
735 |
+
last_hidden_state=sequence_output,
|
736 |
+
pooler_output=pooled_output,
|
737 |
+
)
|
738 |
+
|
739 |
+
|
740 |
+
class NomicBertForPreTraining(NomicBertPreTrainedModel):
|
741 |
+
_tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
|
742 |
+
|
743 |
+
def __init__(self, config: GPT2Config):
|
744 |
+
super().__init__(config)
|
745 |
+
|
746 |
+
self.bert = NomicBertModel(config, add_pooling_layer=getattr(config, "add_pooling_layer", False))
|
747 |
+
self.cls = NomicBertPreTrainingHeads(config)
|
748 |
+
self.mlm_loss = nn.CrossEntropyLoss()
|
749 |
+
|
750 |
+
# Initialize weights and apply final processing
|
751 |
+
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
752 |
+
self.tie_weights()
|
753 |
+
|
754 |
+
def tie_weights(self):
|
755 |
+
self.cls.predictions.decoder.weight = self.bert.embeddings.word_embeddings.weight
|
756 |
+
|
757 |
+
def forward(
|
758 |
+
self,
|
759 |
+
input_ids,
|
760 |
+
position_ids=None,
|
761 |
+
token_type_ids=None,
|
762 |
+
attention_mask=None,
|
763 |
+
labels=None,
|
764 |
+
):
|
765 |
+
"""
|
766 |
+
If labels are provided, they must be -100 for masked out tokens (as specified in the attention
|
767 |
+
mask).
|
768 |
+
Outputs:
|
769 |
+
if `labels` and `next_sentence_label` are not `None`:
|
770 |
+
Outputs the total_loss which is the sum of the masked language modeling loss and the next
|
771 |
+
sentence classification loss.
|
772 |
+
if `labels` or `next_sentence_label` is `None`:
|
773 |
+
Outputs a tuple comprising
|
774 |
+
- the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
|
775 |
+
- the next sentence classification logits of shape [batch_size, 2].
|
776 |
+
|
777 |
+
"""
|
778 |
+
outputs = self.bert(
|
779 |
+
input_ids,
|
780 |
+
position_ids=position_ids,
|
781 |
+
token_type_ids=token_type_ids,
|
782 |
+
attention_mask=attention_mask.bool() if attention_mask is not None else None,
|
783 |
+
)
|
784 |
+
sequence_output, _ = outputs.last_hidden_state, outputs.pooler_output
|
785 |
+
|
786 |
+
prediction_scores = self.cls(sequence_output)
|
787 |
+
|
788 |
+
total_loss = None
|
789 |
+
if labels is not None:
|
790 |
+
masked_lm_loss = self.mlm_loss(
|
791 |
+
rearrange(prediction_scores, "... v -> (...) v"),
|
792 |
+
rearrange(labels, "... -> (...)"),
|
793 |
+
)
|
794 |
+
total_loss = masked_lm_loss.float()
|
795 |
+
|
796 |
+
return BertForPreTrainingOutput(
|
797 |
+
loss=total_loss,
|
798 |
+
prediction_logits=prediction_scores,
|
799 |
+
)
|
800 |
+
|
801 |
+
|
802 |
+
class NomicBertForSequenceClassification(NomicBertPreTrainedModel):
|
803 |
+
def __init__(self, config):
|
804 |
+
super().__init__(config)
|
805 |
+
self.num_labels = config.num_labels
|
806 |
+
self.config = config
|
807 |
+
|
808 |
+
self.bert = NomicBertModel(config)
|
809 |
+
classifier_dropout = (
|
810 |
+
getattr(config, "classifier_dropout", config.embd_pdrop)
|
811 |
+
)
|
812 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
813 |
+
self.classifier = nn.Linear(config.n_embd, config.num_labels)
|
814 |
+
|
815 |
+
# Initialize weights and apply final processing
|
816 |
+
self.post_init()
|
817 |
+
|
818 |
+
def forward(
|
819 |
+
self,
|
820 |
+
input_ids: Optional[torch.Tensor] = None,
|
821 |
+
attention_mask: Optional[torch.Tensor] = None,
|
822 |
+
token_type_ids: Optional[torch.Tensor] = None,
|
823 |
+
position_ids: Optional[torch.Tensor] = None,
|
824 |
+
head_mask: Optional[torch.Tensor] = None,
|
825 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
826 |
+
labels: Optional[torch.Tensor] = None,
|
827 |
+
output_attentions: Optional[bool] = None,
|
828 |
+
output_hidden_states: Optional[bool] = None,
|
829 |
+
return_dict: Optional[bool] = None,
|
830 |
+
):
|
831 |
+
r"""
|
832 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
833 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
834 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
835 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
836 |
+
"""
|
837 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
838 |
+
outputs = self.bert(
|
839 |
+
input_ids,
|
840 |
+
position_ids=position_ids,
|
841 |
+
token_type_ids=token_type_ids,
|
842 |
+
attention_mask=attention_mask.bool() if attention_mask is not None else None,
|
843 |
+
)
|
844 |
+
|
845 |
+
pooled_output = outputs[1]
|
846 |
+
|
847 |
+
pooled_output = self.dropout(pooled_output)
|
848 |
+
logits = self.classifier(pooled_output)
|
849 |
+
|
850 |
+
loss = None
|
851 |
+
if labels is not None:
|
852 |
+
if self.config.problem_type is None:
|
853 |
+
if self.num_labels == 1:
|
854 |
+
self.config.problem_type = "regression"
|
855 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
856 |
+
self.config.problem_type = "single_label_classification"
|
857 |
+
else:
|
858 |
+
self.config.problem_type = "multi_label_classification"
|
859 |
+
|
860 |
+
if self.config.problem_type == "regression":
|
861 |
+
loss_fct = nn.MSELoss()
|
862 |
+
if self.num_labels == 1:
|
863 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
864 |
+
else:
|
865 |
+
loss = loss_fct(logits, labels)
|
866 |
+
elif self.config.problem_type == "single_label_classification":
|
867 |
+
loss_fct = nn.CrossEntropyLoss()
|
868 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
869 |
+
elif self.config.problem_type == "multi_label_classification":
|
870 |
+
loss_fct = nn.BCEWithLogitsLoss()
|
871 |
+
loss = loss_fct(logits, labels)
|
872 |
+
if not return_dict:
|
873 |
+
output = (logits,) + outputs[2:]
|
874 |
+
return ((loss,) + output) if loss is not None else output
|
875 |
+
|
876 |
+
return SequenceClassifierOutput(
|
877 |
+
loss=loss,
|
878 |
+
logits=logits,
|
879 |
+
hidden_states=outputs.hidden_states,
|
880 |
+
attentions=outputs.attentions,
|
881 |
+
)
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e0b1c7db843aeae0c744716acf3c7b7da60d142e7ca31edd11853f5b163c8776
|
3 |
+
size 549328982
|