lv12 commited on
Commit
db39ce8
1 Parent(s): da3d8fe

full set multi loss ESCI triplets

Browse files
1_Pooling/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "word_embedding_dimension": 768,
3
+ "pooling_mode_cls_token": false,
4
+ "pooling_mode_mean_tokens": true,
5
+ "pooling_mode_max_tokens": false,
6
+ "pooling_mode_mean_sqrt_len_tokens": false,
7
+ "pooling_mode_weightedmean_tokens": false,
8
+ "pooling_mode_lasttoken": false,
9
+ "include_prompt": true
10
+ }
README.md ADDED
The diff for this file is too large to render. See raw diff
 
config.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "models/nomic-embed-text-esci/checkpoint-95000",
3
+ "activation_function": "swiglu",
4
+ "architectures": [
5
+ "NomicBertModel"
6
+ ],
7
+ "attn_pdrop": 0.0,
8
+ "auto_map": {
9
+ "AutoConfig": "configuration_hf_nomic_bert.NomicBertConfig",
10
+ "AutoModel": "modeling_hf_nomic_bert.NomicBertModel",
11
+ "AutoModelForMaskedLM": "nomic-ai/nomic-bert-2048--modeling_hf_nomic_bert.NomicBertForPreTraining"
12
+ },
13
+ "bos_token_id": null,
14
+ "causal": false,
15
+ "dense_seq_output": true,
16
+ "embd_pdrop": 0.0,
17
+ "eos_token_id": null,
18
+ "fused_bias_fc": true,
19
+ "fused_dropout_add_ln": true,
20
+ "initializer_range": 0.02,
21
+ "layer_norm_epsilon": 1e-12,
22
+ "max_trained_positions": 2048,
23
+ "mlp_fc1_bias": false,
24
+ "mlp_fc2_bias": false,
25
+ "model_type": "nomic_bert",
26
+ "n_embd": 768,
27
+ "n_head": 12,
28
+ "n_inner": 3072,
29
+ "n_layer": 12,
30
+ "n_positions": 8192,
31
+ "pad_vocab_size_multiple": 64,
32
+ "parallel_block": false,
33
+ "parallel_block_tied_norm": false,
34
+ "prenorm": false,
35
+ "qkv_proj_bias": false,
36
+ "reorder_and_upcast_attn": false,
37
+ "resid_pdrop": 0.0,
38
+ "rotary_emb_base": 1000,
39
+ "rotary_emb_fraction": 1.0,
40
+ "rotary_emb_interleaved": false,
41
+ "rotary_emb_scale_base": null,
42
+ "rotary_scaling_factor": null,
43
+ "scale_attn_by_inverse_layer_idx": false,
44
+ "scale_attn_weights": true,
45
+ "summary_activation": null,
46
+ "summary_first_dropout": 0.0,
47
+ "summary_proj_to_labels": true,
48
+ "summary_type": "cls_index",
49
+ "summary_use_proj": true,
50
+ "torch_dtype": "float32",
51
+ "transformers_version": "4.38.2",
52
+ "type_vocab_size": 2,
53
+ "use_cache": true,
54
+ "use_flash_attn": true,
55
+ "use_rms_norm": false,
56
+ "use_xentropy": true,
57
+ "vocab_size": 30528
58
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "sentence_transformers": "2.4.0.dev0",
4
+ "transformers": "4.37.2",
5
+ "pytorch": "2.1.0+cu121"
6
+ },
7
+ "prompts": {},
8
+ "default_prompt_name": null,
9
+ "similarity_fn_name": null
10
+ }
configuration_hf_nomic_bert.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import GPT2Config
2
+
3
+
4
+ class NomicBertConfig(GPT2Config):
5
+ model_type = "nomic_bert"
6
+
7
+ def __init__(
8
+ self,
9
+ prenorm=False,
10
+ parallel_block=False,
11
+ parallel_block_tied_norm=False,
12
+ rotary_emb_fraction=0.0,
13
+ fused_dropout_add_ln=False,
14
+ fused_bias_fc=False,
15
+ use_flash_attn=False,
16
+ use_xentropy=False,
17
+ qkv_proj_bias=True,
18
+ rotary_emb_base=10_000,
19
+ rotary_emb_scale_base=None,
20
+ rotary_emb_interleaved=False,
21
+ mlp_fc1_bias=True,
22
+ mlp_fc2_bias=True,
23
+ use_rms_norm=False,
24
+ causal=False,
25
+ type_vocab_size=2,
26
+ dense_seq_output=True,
27
+ pad_vocab_size_multiple=1,
28
+ tie_word_embeddings=True,
29
+ rotary_scaling_factor=None,
30
+ max_trained_positions=2048,
31
+ **kwargs,
32
+ ):
33
+ self.prenorm = prenorm
34
+ self.parallel_block = parallel_block
35
+ self.parallel_block_tied_norm = parallel_block_tied_norm
36
+ self.rotary_emb_fraction = rotary_emb_fraction
37
+ self.tie_word_embeddings = tie_word_embeddings
38
+ self.fused_dropout_add_ln = fused_dropout_add_ln
39
+ self.fused_bias_fc = fused_bias_fc
40
+ self.use_flash_attn = use_flash_attn
41
+ self.use_xentropy = use_xentropy
42
+ self.qkv_proj_bias = qkv_proj_bias
43
+ self.rotary_emb_base = rotary_emb_base
44
+ self.rotary_emb_scale_base = rotary_emb_scale_base
45
+ self.rotary_emb_interleaved = rotary_emb_interleaved
46
+ self.mlp_fc1_bias = mlp_fc1_bias
47
+ self.mlp_fc2_bias = mlp_fc2_bias
48
+ self.use_rms_norm = use_rms_norm
49
+ self.causal = causal
50
+ self.type_vocab_size = type_vocab_size
51
+ self.dense_seq_output = dense_seq_output
52
+ self.pad_vocab_size_multiple = pad_vocab_size_multiple
53
+ self.rotary_scaling_factor = rotary_scaling_factor
54
+ self.max_trained_positions = max_trained_positions
55
+
56
+ super().__init__(**kwargs)
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:797bfc53d7dd1fb7ac1de1358a00de3dc0d41bea17004ca8eb999f527d24298a
3
+ size 546938168
modeling_hf_nomic_bert.py ADDED
@@ -0,0 +1,1234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ import logging
7
+
8
+ # Inspired by https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py
9
+ import os
10
+ import re
11
+ from collections import OrderedDict
12
+ from functools import partial
13
+ from typing import List, Optional, Tuple, Union
14
+
15
+ import torch
16
+ import torch.nn as nn
17
+ import torch.nn.functional as F
18
+ from einops import rearrange, repeat
19
+ from safetensors.torch import load_file as safe_load_file
20
+ from transformers import GPT2Config, PreTrainedModel
21
+ from transformers.models.bert.modeling_bert import (
22
+ BaseModelOutputWithPoolingAndCrossAttentions,
23
+ MaskedLMOutput,
24
+ SequenceClassifierOutput,
25
+ )
26
+ from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, WEIGHTS_NAME
27
+ from transformers.utils.hub import cached_file, get_checkpoint_shard_files
28
+
29
+ from .configuration_hf_nomic_bert import NomicBertConfig
30
+
31
+ logger = logging.getLogger(__name__)
32
+
33
+
34
+ # adapted from flash attention, added safe serialization option for hf models
35
+ def state_dict_from_pretrained(model_name, safe_serialization=False, device=None, dtype=None):
36
+ # If not fp32, then we don't want to load directly to the GPU
37
+ mapped_device = "cpu" if dtype not in [torch.float32, None] else device
38
+ is_sharded = False
39
+ load_safe = False
40
+ resolved_archive_file = None
41
+
42
+ weights_path = os.path.join(model_name, WEIGHTS_NAME)
43
+ weights_index_path = os.path.join(model_name, WEIGHTS_INDEX_NAME)
44
+ safe_weights_path = os.path.join(model_name, SAFE_WEIGHTS_NAME)
45
+ safe_weights_index_path = os.path.join(model_name, SAFE_WEIGHTS_INDEX_NAME)
46
+
47
+ if os.path.isfile(weights_path):
48
+ resolved_archive_file = cached_file(model_name, WEIGHTS_NAME, _raise_exceptions_for_missing_entries=False)
49
+ elif os.path.isfile(weights_index_path):
50
+ resolved_archive_file = cached_file(model_name, WEIGHTS_INDEX_NAME, _raise_exceptions_for_missing_entries=False)
51
+ is_sharded = True
52
+ elif os.path.isfile(safe_weights_path):
53
+ resolved_archive_file = cached_file(model_name, SAFE_WEIGHTS_NAME, _raise_exceptions_for_missing_entries=False)
54
+ load_safe = True
55
+ elif os.path.isfile(safe_weights_index_path):
56
+ resolved_archive_file = cached_file(
57
+ model_name, SAFE_WEIGHTS_INDEX_NAME, _raise_exceptions_for_missing_entries=False
58
+ )
59
+ is_sharded = True
60
+ load_safe = True
61
+ else: # Try loading from HF hub instead of from local files
62
+ resolved_archive_file = None
63
+ for weight_name in [WEIGHTS_NAME, SAFE_WEIGHTS_NAME, WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_INDEX_NAME]:
64
+ resolved_archive_file = cached_file(
65
+ model_name, weight_name, _raise_exceptions_for_missing_entries=False
66
+ )
67
+ if resolved_archive_file is not None:
68
+ if weight_name in [SAFE_WEIGHTS_NAME, SAFE_WEIGHTS_INDEX_NAME]:
69
+ load_safe = True
70
+ if weight_name in [WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_INDEX_NAME]:
71
+ is_sharded = True
72
+ break
73
+
74
+ if resolved_archive_file is None:
75
+ raise EnvironmentError(f"Model name {model_name} was not found.")
76
+
77
+ if load_safe:
78
+ loader = partial(safe_load_file, device=mapped_device)
79
+ else:
80
+ loader = partial(torch.load, map_location=mapped_device)
81
+
82
+ if is_sharded:
83
+ # resolved_archive_file becomes a list of files that point to the different
84
+ # checkpoint shards in this case.
85
+ resolved_archive_file, sharded_metadata = get_checkpoint_shard_files(model_name, resolved_archive_file)
86
+ state_dict = {}
87
+ for sharded_file in resolved_archive_file:
88
+ state_dict.update(loader(sharded_file))
89
+ else:
90
+ state_dict = loader(resolved_archive_file)
91
+ # Convert dtype before moving to GPU to save memory
92
+ if dtype is not None:
93
+ state_dict = {k: v.to(dtype=dtype) for k, v in state_dict.items()}
94
+ state_dict = {k: v.to(device=device) for k, v in state_dict.items()}
95
+ return state_dict
96
+
97
+
98
+ def filter_shapes(state_dict, model):
99
+ """
100
+ Filters the state dict to match the current model shape.
101
+ """
102
+ filtered_state_dict = {}
103
+ for key, value in state_dict.items():
104
+ if key in model.state_dict():
105
+ if value.shape == model.state_dict()[key].shape:
106
+ filtered_state_dict[key] = value
107
+ return filtered_state_dict
108
+
109
+
110
+ def remap_bert_state_dict(
111
+ state_dict,
112
+ config,
113
+ remove_bert=False,
114
+ remove_cls_weights=False,
115
+ add_pooling_layer=False,
116
+ ):
117
+ """
118
+ Map the state_dict of a Huggingface BERT model to be flash_attn compatible.
119
+ """
120
+
121
+ def add_bert_prefix(key):
122
+ # prepend bert. to the key
123
+ if key.startswith("bert.") or key.startswith("cls."):
124
+ return key
125
+ return f"bert.{key}"
126
+
127
+ state_dict = OrderedDict((add_bert_prefix(k), v) for k, v in state_dict.items())
128
+
129
+ # LayerNorm
130
+ def key_mapping_ln_gamma_beta(key):
131
+ key = re.sub(r"LayerNorm.gamma$", "LayerNorm.weight", key)
132
+ key = re.sub(r"LayerNorm.beta$", "LayerNorm.bias", key)
133
+ return key
134
+
135
+ state_dict = OrderedDict((key_mapping_ln_gamma_beta(k), v) for k, v in state_dict.items())
136
+
137
+ # Layers
138
+ def key_mapping_layers(key):
139
+ return re.sub(r"^bert.encoder.layer\.", "bert.encoder.layers.", key)
140
+
141
+ state_dict = OrderedDict((key_mapping_layers(k), v) for k, v in state_dict.items())
142
+
143
+ # LayerNorm
144
+ def key_mapping_ln(key):
145
+ key = re.sub(r"^bert.embeddings.LayerNorm.", "bert.emb_ln.", key)
146
+ key = re.sub(
147
+ r"^bert.encoder.layers.(\d+).attention.output.LayerNorm.(weight|bias)",
148
+ r"bert.encoder.layers.\1.norm1.\2",
149
+ key,
150
+ )
151
+ key = re.sub(
152
+ r"^bert.encoder.layers.(\d+).output.LayerNorm.(weight|bias)",
153
+ r"bert.encoder.layers.\1.norm2.\2",
154
+ key,
155
+ )
156
+ key = re.sub(
157
+ r"^cls.predictions.transform.LayerNorm.(weight|bias)",
158
+ r"cls.predictions.transform.layer_norm.\1",
159
+ key,
160
+ )
161
+ return key
162
+
163
+ state_dict = OrderedDict((key_mapping_ln(k), v) for k, v in state_dict.items())
164
+
165
+ # MLP
166
+ def key_mapping_mlp(key):
167
+ key = re.sub(
168
+ r"^bert.encoder.layers.(\d+).intermediate.dense.(weight|bias)",
169
+ r"bert.encoder.layers.\1.mlp.fc1.\2",
170
+ key,
171
+ )
172
+ key = re.sub(
173
+ r"^bert.encoder.layers.(\d+).output.dense.(weight|bias)",
174
+ r"bert.encoder.layers.\1.mlp.fc2.\2",
175
+ key,
176
+ )
177
+ return key
178
+
179
+ state_dict = OrderedDict((key_mapping_mlp(k), v) for k, v in state_dict.items())
180
+
181
+ # Attention
182
+ last_layer_subset = getattr(config, "last_layer_subset", False)
183
+ for d in range(config.num_hidden_layers):
184
+ if f"bert.encoder.layers.{d}.attention.self.query.weight" not in state_dict:
185
+ continue
186
+ Wq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.weight")
187
+ Wk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.weight")
188
+ Wv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.weight")
189
+ bq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.bias")
190
+ bk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.bias")
191
+ bv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.bias")
192
+ if not (last_layer_subset and d == config.num_hidden_layers - 1):
193
+ state_dict[f"bert.encoder.layers.{d}.attn.Wqkv.weight"] = torch.cat([Wq, Wk, Wv], dim=0)
194
+ state_dict[f"bert.encoder.layers.{d}.attn.Wqkv.bias"] = torch.cat([bq, bk, bv], dim=0)
195
+ else:
196
+ state_dict[f"bert.encoder.layers.{d}.attn.Wq.weight"] = Wq
197
+ state_dict[f"bert.encoder.layers.{d}.attn.Wkv.weight"] = torch.cat([Wk, Wv], dim=0)
198
+ state_dict[f"bert.encoder.layers.{d}.attn.Wq.bias"] = bq
199
+ state_dict[f"bert.encoder.layers.{d}.attn.Wkv.bias"] = torch.cat([bk, bv], dim=0)
200
+
201
+ def key_mapping_attn(key):
202
+ return re.sub(
203
+ r"^bert.encoder.layers.(\d+).attention.output.dense.(weight|bias)",
204
+ r"bert.encoder.layers.\1.attn.out_proj.\2",
205
+ key,
206
+ )
207
+
208
+ state_dict = OrderedDict((key_mapping_attn(k), v) for k, v in state_dict.items())
209
+
210
+ def key_mapping_decoder_bias(key):
211
+ return re.sub(r"^cls.predictions.bias", "cls.predictions.decoder.bias", key)
212
+
213
+ # remove nsp weights, we don't use
214
+ state_dict.pop("cls.seq_relationship.weight", None)
215
+ state_dict.pop("cls.seq_relationship.bias", None)
216
+ state_dict.pop("bert.embeddings.position_ids", None)
217
+
218
+ state_dict = OrderedDict((key_mapping_decoder_bias(k), v) for k, v in state_dict.items())
219
+
220
+ if remove_cls_weights:
221
+ cls_weights = [
222
+ "cls.predictions.decoder.bias",
223
+ "cls.predictions.transform.dense.weight",
224
+ "cls.predictions.transform.dense.bias",
225
+ "cls.predictions.transform.layer_norm.weight",
226
+ "cls.predictions.transform.layer_norm.bias",
227
+ "cls.predictions.decoder.weight",
228
+ ]
229
+ for weight in cls_weights:
230
+ state_dict.pop(weight, None)
231
+
232
+ # Word embedding
233
+ pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
234
+ if pad_vocab_size_multiple > 1:
235
+ word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
236
+ state_dict["bert.embeddings.word_embeddings.weight"] = F.pad(
237
+ word_embeddings, (0, 0, 0, config.vocab_size - word_embeddings.shape[0])
238
+ )
239
+ if not remove_cls_weights:
240
+ decoder_weight = state_dict["cls.predictions.decoder.weight"]
241
+ state_dict["cls.predictions.decoder.weight"] = F.pad(
242
+ decoder_weight, (0, 0, 0, config.vocab_size - decoder_weight.shape[0])
243
+ )
244
+ # If the vocab was padded, we want to set the decoder bias for those padded indices to be
245
+ # strongly negative (i.e. the decoder shouldn't predict those indices).
246
+ # TD [2022-05-09]: I don't think it affects the MLPerf training.
247
+ if "cls.predictions.decoder.bias" in state_dict:
248
+ decoder_bias = state_dict["cls.predictions.decoder.bias"]
249
+ state_dict["cls.predictions.decoder.bias"] = F.pad(
250
+ decoder_bias, (0, config.vocab_size - decoder_bias.shape[0]), value=-100.0
251
+ )
252
+
253
+ if add_pooling_layer is False:
254
+ pooler_weights = [
255
+ "bert.pooler.dense.weight",
256
+ "bert.pooler.dense.bias",
257
+ ]
258
+ for key in pooler_weights:
259
+ state_dict.pop(key, None)
260
+
261
+ if remove_bert:
262
+
263
+ def remove_bert_prefix(key):
264
+ key = re.sub(r"^bert.", "", key)
265
+ return key
266
+
267
+ state_dict = OrderedDict((remove_bert_prefix(k), v) for k, v in state_dict.items())
268
+
269
+ return state_dict
270
+
271
+
272
+ class NomicBertPreTrainedModel(PreTrainedModel):
273
+ """An abstract class to handle weights initialization and
274
+ a simple interface for dowloading and loading pretrained models.
275
+ """
276
+
277
+ config_class = NomicBertConfig
278
+ base_model_prefix = "model"
279
+ supports_gradient_checkpointing = True
280
+ _no_split_modules = ["Block"]
281
+ _skip_keys_device_placement = "past_key_values"
282
+
283
+ def __init__(self, config, *inputs, **kwargs):
284
+ super().__init__(config)
285
+ if not isinstance(config, GPT2Config):
286
+ raise ValueError(
287
+ "Parameter config in `{}(config)` should be an instance of class `GPT2Config`. "
288
+ "To create a model from a Google pretrained model use "
289
+ "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
290
+ self.__class__.__name__, self.__class__.__name__
291
+ )
292
+ )
293
+ self.config = config
294
+
295
+ @classmethod
296
+ def from_pretrained(cls, model_name, config=None, *inputs, **kwargs):
297
+ """
298
+ Instantiate a NomicBertPreTrainedModel from a pre-trained model file or a pytorch state dict.
299
+ Download and cache the pre-trained model file if needed.
300
+
301
+ Params:
302
+ pretrained_model_name_or_path: either:
303
+ - a path or url to a pretrained model archive containing:
304
+ . `bert_config.json` a configuration file for the model
305
+ . `pytorch_model.bin` a PyTorch dump of a NomicBertForPretraining instance
306
+ - a path or url to a pretrained model archive containing:
307
+ . `bert_config.json` a configuration file for the model
308
+ . `model.chkpt` a TensorFlow checkpoint
309
+ *inputs, **kwargs: additional input for the specific NomicBert class
310
+ (ex: num_labels for NomicBertForSequenceClassification)
311
+ """
312
+ # Instantiate model.
313
+ if config is None:
314
+ config = cls.config_class.from_pretrained(model_name)
315
+ remove_cls = cls != NomicBertForPreTraining
316
+ remove_bert_prefix = cls != NomicBertForPreTraining and cls != NomicBertForSequenceClassification
317
+ ignore_mismatched_shapes = kwargs.pop("ignore_mismatched_sizes", False)
318
+ num_labels = kwargs.pop("num_labels", None)
319
+ rotary_scaling_factor = kwargs.pop("rotary_scaling_factor", None)
320
+ strict = kwargs.pop("strict", True)
321
+ if rotary_scaling_factor:
322
+ config.rotary_scaling_factor = rotary_scaling_factor
323
+
324
+ if config.n_positions <= 0 and config.rotary_emb_fraction > 0:
325
+ config.n_positions = 2048
326
+ if num_labels:
327
+ config.num_labels = num_labels
328
+
329
+ if "add_pooling_layer" in kwargs:
330
+ model = cls(config, *inputs, add_pooling_layer=kwargs.pop("add_pooling_layer"))
331
+ else:
332
+ if cls == NomicBertModel:
333
+ model = cls(config, *inputs, add_pooling_layer=False)
334
+ else:
335
+ model = cls(config, *inputs)
336
+ # TODO: fix this
337
+ # Assuming we know what we're doing when loading from disk
338
+ # Prob a bad assumption but i'm tired and want to train this asap
339
+ if os.path.exists(model_name):
340
+ model_path = f"{model_name}/pytorch_model.bin"
341
+ if os.path.exists(model_path):
342
+ state_dict = torch.load(f"{model_name}/pytorch_model.bin")
343
+ else:
344
+ model_path = f"{model_name}/model.safetensors"
345
+ if not os.path.exists(model_path):
346
+ raise ValueError(f"Model path {model_path} not found")
347
+ state_dict = safe_load_file(model_path)
348
+
349
+ if ignore_mismatched_shapes:
350
+ state_dict = filter_shapes(state_dict, model)
351
+ load_return = model.load_state_dict(state_dict, strict=False)
352
+ else:
353
+ # TODO: can probably check config class and see if we need to remap from a bert model
354
+ state_dict = state_dict_from_pretrained(model_name)
355
+ state_dict = remap_bert_state_dict(
356
+ state_dict,
357
+ config,
358
+ remove_bert=remove_bert_prefix,
359
+ remove_cls_weights=remove_cls,
360
+ add_pooling_layer=getattr(config, "add_pooling_layer", False),
361
+ )
362
+ if ignore_mismatched_shapes:
363
+ state_dict = filter_shapes(state_dict, model)
364
+
365
+ load_return = model.load_state_dict(state_dict, strict=strict)
366
+ logger.warning(load_return)
367
+ return model
368
+
369
+ def _set_gradient_checkpointing(self, module, value=False):
370
+ if isinstance(module, NomicBertEncoder):
371
+ module.gradient_checkpointing = value
372
+
373
+
374
+ # https://github.com/huggingface/transformers/blob/7032e0203262ebb2ebf55da8d2e01f873973e835/src/transformers/models/bert/modeling_bert.py#L748
375
+ def _init_weights(module, initializer_range=0.02):
376
+ if isinstance(module, nn.Linear):
377
+ nn.init.normal_(module.weight, std=initializer_range)
378
+ if module.bias is not None:
379
+ nn.init.zeros_(module.bias)
380
+ elif isinstance(module, nn.Embedding):
381
+ nn.init.normal_(module.weight, std=initializer_range)
382
+ if module.padding_idx is not None:
383
+ nn.init.zeros_(module.weight[module.padding_idx])
384
+
385
+
386
+ class NomicBertEmbeddings(nn.Module):
387
+ def __init__(self, config):
388
+ """
389
+ If max_position_embeddings <= 0, there's no position embeddings
390
+ If type_vocab_size <= 0, there's no token type embeddings
391
+ """
392
+ super().__init__()
393
+ self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
394
+ self.max_position_embeddings = config.max_position_embeddings if config.rotary_emb_fraction <= 0 else 0
395
+ self.type_vocab_size = config.type_vocab_size
396
+ if self.max_position_embeddings > 0 and config.rotary_emb_fraction <= 0:
397
+ self.position_embeddings = nn.Embedding(
398
+ config.max_position_embeddings,
399
+ config.hidden_size,
400
+ )
401
+ if self.type_vocab_size > 0:
402
+ self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
403
+
404
+ def forward(self, input_ids, position_ids=None, token_type_ids=None):
405
+ """
406
+ input_ids: (batch, seqlen)
407
+ position_ids: (batch, seqlen)
408
+ token_type_ids: (batch, seqlen)
409
+ """
410
+ batch_size, seqlen = input_ids.shape
411
+ embeddings = self.word_embeddings(input_ids)
412
+
413
+ if self.type_vocab_size > 0:
414
+ if token_type_ids is None:
415
+ token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
416
+ token_type_embeddings = self.token_type_embeddings(token_type_ids)
417
+ embeddings = embeddings + token_type_embeddings
418
+
419
+ if self.max_position_embeddings > 0:
420
+ if position_ids is None:
421
+ position_ids = torch.arange(seqlen, dtype=torch.long, device=input_ids.device)
422
+ position_embeddings = self.position_embeddings(position_ids)
423
+ embeddings = embeddings + position_embeddings
424
+ return embeddings
425
+
426
+
427
+ class NomicBertMLP(nn.Module):
428
+ def __init__(
429
+ self,
430
+ in_features,
431
+ hidden_features=None,
432
+ out_features=None,
433
+ activation=F.gelu,
434
+ bias1=True,
435
+ bias2=True,
436
+ return_residual=False,
437
+ fused_bias_fc=False,
438
+ ):
439
+ super().__init__()
440
+ out_features = out_features if out_features is not None else in_features
441
+ hidden_features = hidden_features if hidden_features is not None else in_features * 4
442
+ self.return_residual = return_residual
443
+ self.fc1 = nn.Linear(in_features, hidden_features, bias=bias1)
444
+ approximate = "tanh" if activation in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"] else "none"
445
+ self.activation = nn.GELU(approximate=approximate) if activation == "gelu" else activation
446
+ self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2)
447
+
448
+ def forward(self, x):
449
+ y = self.fc1(x)
450
+ y = self.activation(y)
451
+ y = self.fc2(y)
452
+ return y if not self.return_residual else (y, x)
453
+
454
+
455
+ class NomciBertGatedMLP(nn.Module):
456
+ def __init__(
457
+ self,
458
+ in_features,
459
+ hidden_features=None,
460
+ out_features=None,
461
+ activation=F.sigmoid,
462
+ bias1=True,
463
+ bias2=True,
464
+ multiple_of=256,
465
+ return_residual=False,
466
+ fused_bias_fc=True,
467
+ device=None,
468
+ dtype=None,
469
+ ):
470
+ super().__init__()
471
+ out_features = out_features if out_features is not None else in_features
472
+ hidden_features = hidden_features if hidden_features is not None else int(8 * in_features / 3)
473
+ hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
474
+ self.return_residual = return_residual
475
+
476
+ self.fc11 = nn.Linear(in_features, hidden_features, bias=bias1)
477
+ self.fc12 = nn.Linear(in_features, hidden_features, bias=bias1)
478
+ self.activation = activation
479
+ self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2)
480
+
481
+ def forward(self, x):
482
+ y = self.fc11(x)
483
+ gate = self.fc12(x)
484
+ if self.activation == F.sigmoid: # Special case for GLU
485
+ y = F.glu(torch.cat([y, gate], dim=-1), dim=-1)
486
+ else:
487
+ y = y * self.activation(gate)
488
+ y = self.fc2(y)
489
+ return y if not self.return_residual else (y, x)
490
+
491
+
492
+ def rotate_half(x, interleaved=False):
493
+ if not interleaved:
494
+ x1, x2 = x.chunk(2, dim=-1)
495
+ return torch.cat((-x2, x1), dim=-1)
496
+ else:
497
+ x1, x2 = x[..., ::2], x[..., 1::2]
498
+ return rearrange(torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2)
499
+
500
+
501
+ def apply_rotary_emb(x, cos, sin, offset=0, interleaved=False):
502
+ """
503
+ x: (batch_size, seqlen, nheads, headdim)
504
+ cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
505
+ """
506
+ ro_dim = cos.shape[-1] * 2
507
+ assert ro_dim <= x.shape[-1]
508
+ cos, sin = (
509
+ cos[offset : offset + x.shape[1]],
510
+ sin[offset : offset + x.shape[1]],
511
+ )
512
+ cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
513
+ sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
514
+ return torch.cat(
515
+ [x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]],
516
+ dim=-1,
517
+ )
518
+
519
+
520
+ class NomicBertRotaryEmbedding(nn.Module):
521
+ def __init__(
522
+ self,
523
+ dim: int,
524
+ base=10000.0,
525
+ interleaved=False,
526
+ scale_base=None,
527
+ pos_idx_in_fp32=True,
528
+ device=None,
529
+ ):
530
+ """
531
+ interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
532
+ of 1st half and 2nd half (GPT-NeoX style).
533
+ pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
534
+ otherwise they might be in lower precision.
535
+ This option was added because previously (before 2023-07-02), when we construct
536
+ the position indices, we use the dtype of self.inv_freq. In most cases this would
537
+ be fp32, but if the model is trained in pure bf16 (not mixed precision), then
538
+ self.inv_freq would be bf16, and the position indices are also in bf16.
539
+ Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
540
+ embeddings for some positions will coincide.
541
+ To maintain compatibility with models previously trained in pure bf16,
542
+ we add this option.
543
+ """
544
+ super().__init__()
545
+ self.dim = dim
546
+ self.base = float(base)
547
+ self.pos_idx_in_fp32 = pos_idx_in_fp32
548
+ # Generate and save the inverse frequency buffer (non trainable)
549
+ inv_freq = self._compute_inv_freq(device)
550
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
551
+ self.interleaved = interleaved
552
+ self.scale_base = scale_base
553
+ scale = (
554
+ (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
555
+ if scale_base is not None
556
+ else None
557
+ )
558
+ self.register_buffer("scale", scale, persistent=False)
559
+
560
+ self._seq_len_cached = 0
561
+ self._cos_cached = None
562
+ self._sin_cached = None
563
+ self._cos_k_cached = None
564
+ self._sin_k_cached = None
565
+
566
+ def _compute_inv_freq(self, device=None):
567
+ return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
568
+
569
+ def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
570
+ # Reset the tables if the sequence length has changed,
571
+ # if we're on a new device (possibly due to tracing for instance),
572
+ # or if we're switching from inference mode to training
573
+ if (
574
+ seqlen > self._seq_len_cached
575
+ or self._cos_cached is None
576
+ or self._cos_cached.device != device
577
+ or self._cos_cached.dtype != dtype
578
+ or (self.training and self._cos_cached.is_inference())
579
+ ):
580
+ self._seq_len_cached = seqlen
581
+ # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
582
+ # And the output of arange can be quite large, so bf16 would lose a lot of precision.
583
+ # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
584
+ if self.pos_idx_in_fp32:
585
+ t = torch.arange(seqlen, device=device, dtype=torch.float32)
586
+ # We want fp32 here as well since inv_freq will be multiplied with t, and the output
587
+ # will be large. Having it in bf16 will lose a lot of precision and cause the
588
+ # cos & sin output to change significantly.
589
+ # We want to recompute self.inv_freq if it was not loaded in fp32
590
+ if self.inv_freq.dtype != torch.float32:
591
+ inv_freq = self._compute_inv_freq(device=device)
592
+ else:
593
+ inv_freq = self.inv_freq
594
+ else:
595
+ t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
596
+ inv_freq = self.inv_freq
597
+ # Don't do einsum, it converts fp32 to fp16 under AMP
598
+ # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
599
+ freqs = torch.outer(t, inv_freq)
600
+ self._cos_cached = torch.cos(freqs).to(dtype)
601
+ self._sin_cached = torch.sin(freqs).to(dtype)
602
+
603
+ def forward(
604
+ self,
605
+ qkv: torch.Tensor,
606
+ kv: Optional[torch.Tensor] = None,
607
+ seqlen_offset: Union[int, torch.Tensor] = 0,
608
+ max_seqlen: Optional[int] = None,
609
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
610
+ """
611
+ qkv: (batch, seqlen, 3, nheads, headdim) if kv is none,
612
+ else it's just q of shape (batch, seqlen, nheads, headdim)
613
+ kv: (batch, seqlen, 2, nheads, headdim)
614
+ seqlen_offset: (batch_size,) or int. Each sequence in x is shifted by this amount.
615
+ Most commonly used in inference when we have KV cache.
616
+ If it's a tensor of shape (batch_size,), then to update the cos / sin cache, one
617
+ should pass in max_seqlen, which will update the cos / sin cache up to that length.
618
+ Apply rotary embedding *inplace* to qkv and / or kv.
619
+ """
620
+ seqlen = qkv.shape[1]
621
+ if seqlen > self._seq_len_cached:
622
+ self._update_cos_sin_cache(seqlen, device=qkv.device, dtype=qkv.dtype)
623
+ elif max_seqlen is not None:
624
+ self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
625
+ elif isinstance(seqlen_offset, int):
626
+ self._update_cos_sin_cache(seqlen + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
627
+
628
+ q_rot = apply_rotary_emb(qkv[:, :, 0], self._cos_cached, self._sin_cached, seqlen_offset, self.interleaved)
629
+ k_rot = apply_rotary_emb(qkv[:, :, 1], self._cos_cached, self._sin_cached, seqlen_offset, self.interleaved)
630
+ return torch.stack((q_rot, k_rot, qkv[:, :, 2]), dim=2)
631
+
632
+
633
+ class NomicBertDynamicNTKRotaryEmbedding(NomicBertRotaryEmbedding):
634
+ def __init__(self, rotary_scaling_factor, max_position_embeddings, **kwargs):
635
+ super().__init__(**kwargs)
636
+ self.rotary_scaling_factor = rotary_scaling_factor
637
+ self.max_position_embeddings = max_position_embeddings
638
+
639
+ def _compute_inv_freq(self, base=None, device=None):
640
+ if base is None:
641
+ base = self.base
642
+ return 1.0 / (base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
643
+
644
+ def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
645
+ # Reset the tables if the sequence length has changed,
646
+ # if we're on a new device (possibly due to tracing for instance),
647
+ # or if we're switching from inference mode to training
648
+ if seqlen > self.max_position_embeddings:
649
+ base = self.base * (
650
+ (self.rotary_scaling_factor * seqlen / self.max_position_embeddings) - (self.rotary_scaling_factor - 1)
651
+ ) ** (self.dim / (self.dim - 2))
652
+ inv_freq = self._compute_inv_freq(base=base, device=device)
653
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
654
+
655
+ if (
656
+ seqlen > self._seq_len_cached
657
+ or self._cos_cached is None
658
+ or self._cos_cached.device != device
659
+ or self._cos_cached.dtype != dtype
660
+ or (self.training and self._cos_cached.is_inference())
661
+ ):
662
+ self._seq_len_cached = seqlen
663
+ # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
664
+ # And the output of arange can be quite large, so bf16 would lose a lot of precision.
665
+ # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
666
+ if self.pos_idx_in_fp32:
667
+ t = torch.arange(seqlen, device=device, dtype=torch.float32)
668
+ # We want fp32 here as well since inv_freq will be multiplied with t, and the output
669
+ # will be large. Having it in bf16 will lose a lot of precision and cause the
670
+ # cos & sin output to change significantly.
671
+ # We want to recompute self.inv_freq if it was not loaded in fp32
672
+ if self.inv_freq.dtype != torch.float32:
673
+ if seqlen > self.max_position_embeddings:
674
+ base = self.base * (
675
+ (self.scaling_factor * seqlen / self.max_position_embeddings) - (self.scaling_factor - 1)
676
+ ) ** (self.dim / (self.dim - 2))
677
+ else:
678
+ base = self.base
679
+ inv_freq = self._compute_inv_freq(device=device, base=base)
680
+ else:
681
+ inv_freq = self.inv_freq
682
+ else:
683
+ t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
684
+ inv_freq = self.inv_freq
685
+ # Don't do einsum, it converts fp32 to fp16 under AMP
686
+ # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
687
+ freqs = torch.outer(t, inv_freq)
688
+ if self.scale is None:
689
+ self._cos_cached = torch.cos(freqs).to(dtype)
690
+ self._sin_cached = torch.sin(freqs).to(dtype)
691
+ else:
692
+ power = (
693
+ torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
694
+ ) / self.scale_base
695
+ scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
696
+ # We want the multiplication by scale to happen in fp32
697
+ self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
698
+ self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
699
+ self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
700
+ self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
701
+
702
+
703
+ class NomicBertAttention(nn.Module):
704
+ """Multi-head self-attention and cross-attention"""
705
+
706
+ def __init__(
707
+ self,
708
+ config,
709
+ ) -> None:
710
+ """
711
+ num_heads_kv: can be used to toggle MQA / GQA. If None, use num_heads.
712
+ return_residual: whether to return the input x along with the output. This is for
713
+ performance reason: for post-norm architecture, returning the input allows us
714
+ to fuse the backward of nn.Linear with the residual connection.
715
+ """
716
+ super().__init__()
717
+ self.embed_dim = config.n_embd
718
+ self.use_flash_attn = config.use_flash_attn
719
+ self.fused_bias_fc = config.fused_bias_fc
720
+
721
+ self.num_heads = config.n_head
722
+ self.num_heads_kv = config.num_heads_kv if getattr(config, "num_heads_kv", None) is not None else self.num_heads
723
+ assert self.embed_dim % self.num_heads == 0, "embed_dim must be divisible by num_heads"
724
+ self.head_dim = self.embed_dim // self.num_heads
725
+ # we don't really support mqa / gqa for now
726
+ qkv_dim = self.head_dim * (self.num_heads + 2 * self.num_heads_kv)
727
+
728
+ self.register_buffer(
729
+ "norm_factor",
730
+ torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()),
731
+ persistent=False,
732
+ )
733
+
734
+ self.rotary_emb_dim = self.head_dim * config.rotary_emb_fraction
735
+ if self.rotary_emb_dim > 0:
736
+ if getattr(config, "rotary_scaling_factor", None):
737
+ self.rotary_emb = NomicBertDynamicNTKRotaryEmbedding(
738
+ dim=self.rotary_emb_dim,
739
+ base=config.rotary_emb_base,
740
+ scale_base=config.rotary_emb_scale_base,
741
+ interleaved=config.rotary_emb_interleaved,
742
+ rotary_scaling_factor=config.rotary_scaling_factor,
743
+ max_position_embeddings=config.max_trained_positions,
744
+ )
745
+ else:
746
+ self.rotary_emb = NomicBertRotaryEmbedding(
747
+ dim=self.rotary_emb_dim,
748
+ base=config.rotary_emb_base,
749
+ scale_base=config.rotary_emb_scale_base,
750
+ interleaved=config.rotary_emb_interleaved,
751
+ )
752
+ # bug in xformers: https://github.com/facebookresearch/xformers/issues/841
753
+ # uses the head dimension instead of the sequence dimension
754
+ self.rotary_head_dim = getattr(config, "rotary_head_dim", False)
755
+
756
+ self.Wqkv = nn.Linear(self.embed_dim, qkv_dim, bias=config.qkv_proj_bias)
757
+
758
+ self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.qkv_proj_bias)
759
+ self.causal = config.causal
760
+ self.drop = nn.Dropout(config.attn_pdrop)
761
+
762
+ def forward(
763
+ self,
764
+ hidden_states: torch.Tensor,
765
+ attention_mask: Optional[torch.Tensor] = None,
766
+ position_ids: Optional[torch.LongTensor] = None,
767
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
768
+ output_attentions: bool = False,
769
+ use_cache: bool = False,
770
+ is_padded_inputs: Optional[bool] = True,
771
+ cu_seqlens: Optional[torch.Tensor] = None,
772
+ max_seq_len: Optional[int] = None,
773
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
774
+
775
+ has_layer_past = past_key_value is not None
776
+
777
+ if has_layer_past:
778
+ past_key_value = past_key_value[0]
779
+ past_len = past_key_value[1]
780
+ else:
781
+ past_len = 0
782
+
783
+ qkv = self.Wqkv(hidden_states)
784
+ qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
785
+
786
+ past_key_value = (past_key_value, past_len + qkv.size(1)) if use_cache else None
787
+
788
+ if self.rotary_emb_dim > 0:
789
+ if self.rotary_head_dim:
790
+ qkv = rearrange(qkv, "b s three h d -> b h three s d")
791
+ qkv = self.rotary_emb(qkv, seqlen_offset=past_len)
792
+
793
+ if self.rotary_head_dim:
794
+ qkv = rearrange(qkv, "b h three s d -> b s three h d")
795
+
796
+ query, key, value = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2]
797
+
798
+ query = query.permute(0, 2, 1, 3)
799
+ key = key.permute(0, 2, 1, 3)
800
+ value = value.permute(0, 2, 1, 3)
801
+
802
+ attention_scores = torch.matmul(query, key.transpose(-1, -2)) / self.norm_factor
803
+ if attention_mask is not None:
804
+ attention_scores = attention_scores + attention_mask
805
+
806
+ attentions_probs = F.softmax(attention_scores, dim=-1)
807
+ attentions_probs = self.drop(attentions_probs)
808
+
809
+ attn_output = torch.matmul(attentions_probs, value)
810
+ attn_output = rearrange(attn_output.permute(0, 2, 1, 3), "... h d -> ... (h d)")
811
+
812
+ attn_output = self.out_proj(attn_output)
813
+
814
+ return attn_output
815
+
816
+
817
+ class NomicBertBlock(NomicBertPreTrainedModel):
818
+ def __init__(
819
+ self,
820
+ config,
821
+ ):
822
+ super().__init__(config=config)
823
+ self.prenorm = config.prenorm
824
+ self.fused_dropout_add_ln = config.fused_dropout_add_ln
825
+
826
+ self.attn = NomicBertAttention(config)
827
+ activation = (
828
+ F.sigmoid
829
+ if config.activation_function == "glu"
830
+ else (F.silu if config.activation_function == "swiglu" else F.gelu)
831
+ )
832
+ if config.activation_function in ["glu", "swiglu", "geglu"]:
833
+ self.mlp = NomciBertGatedMLP(
834
+ config.n_embd,
835
+ hidden_features=config.n_inner,
836
+ bias1=config.mlp_fc1_bias,
837
+ bias2=config.mlp_fc2_bias,
838
+ activation=activation,
839
+ fused_bias_fc=config.fused_bias_fc,
840
+ )
841
+ else:
842
+ self.mlp = NomicBertMLP(
843
+ config.n_embd,
844
+ hidden_features=config.n_inner,
845
+ bias1=config.mlp_fc1_bias,
846
+ bias2=config.mlp_fc2_bias,
847
+ activation=activation,
848
+ fused_bias_fc=config.fused_bias_fc,
849
+ )
850
+
851
+ self.dropout1 = nn.Dropout(config.resid_pdrop)
852
+ self.norm1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
853
+ self.norm2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
854
+ self.dropout2 = nn.Dropout(config.resid_pdrop)
855
+
856
+ def forward(
857
+ self,
858
+ hidden_states: torch.Tensor,
859
+ hidden_states2: torch.Tensor,
860
+ residual: Optional[torch.Tensor] = None,
861
+ attention_mask: Optional[torch.Tensor] = None,
862
+ position_ids: Optional[torch.LongTensor] = None,
863
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
864
+ is_padded_inputs: Optional[bool] = True,
865
+ output_attentions: Optional[bool] = False,
866
+ use_cache: Optional[bool] = False,
867
+ cu_seqlens: Optional[torch.Tensor] = None,
868
+ max_seq_len: Optional[int] = None,
869
+ ):
870
+ r"""Pass the input through the encoder layer.
871
+
872
+ Args:
873
+ hidden_states: the sequence to the encoder layer (required).
874
+ residual: if postnorm, residual=None, If prenorm, hidden_states = Attn/MLP(LN(residual))
875
+ mixer_subset: for cross-attention only. If not None, will take a subset of x
876
+ before applying the query projection. Useful for e.g., ViT where we only care
877
+ about the CLS token in the last layer.
878
+ """
879
+ if self.prenorm:
880
+ dropped = self.dropout1(hidden_states)
881
+ residual = (dropped + residual) if residual is not None else dropped
882
+ hidden_states = self.norm1(residual.to(dtype=self.norm1.weight.dtype))
883
+ hidden_states = self.attn(
884
+ hidden_states,
885
+ attention_mask=attention_mask,
886
+ is_padded_inputs=is_padded_inputs,
887
+ cu_seqlens=cu_seqlens,
888
+ max_seq_len=max_seq_len,
889
+ )
890
+
891
+ dropped = self.dropout2(hidden_states)
892
+ residual = (dropped + residual) if residual is not None else dropped
893
+ hidden_states = self.norm2(residual.to(dtype=self.norm2.weight.dtype))
894
+ hidden_states = self.mlp(hidden_states)
895
+
896
+ return hidden_states, None, residual
897
+ else:
898
+ assert residual is None
899
+ attn_outputs = self.attn(
900
+ hidden_states,
901
+ attention_mask=attention_mask,
902
+ is_padded_inputs=is_padded_inputs,
903
+ cu_seqlens=cu_seqlens,
904
+ max_seq_len=max_seq_len,
905
+ )
906
+ hidden_states = self.norm1((self.dropout1(attn_outputs) + hidden_states).to(dtype=self.norm1.weight.dtype))
907
+ mlp_out = self.mlp(hidden_states)
908
+
909
+ hidden_states = self.norm2((self.dropout2(mlp_out) + hidden_states).to(dtype=self.norm2.weight.dtype))
910
+ return hidden_states, None, None
911
+
912
+
913
+ class NomicBertEncoder(nn.Module):
914
+ def __init__(self, config: GPT2Config):
915
+ super().__init__()
916
+ self.layers = nn.ModuleList([NomicBertBlock(config) for _ in range(config.n_layer)])
917
+ self.gradient_checkpointing = False
918
+ self.config = config
919
+
920
+ def forward(
921
+ self,
922
+ hidden_states: torch.LongTensor = None,
923
+ attention_mask: Optional[torch.Tensor] = None,
924
+ position_ids: Optional[torch.LongTensor] = None,
925
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
926
+ inputs_embeds: Optional[torch.FloatTensor] = None,
927
+ use_cache: Optional[bool] = None,
928
+ output_attentions: Optional[bool] = None,
929
+ output_hidden_states: Optional[bool] = None,
930
+ return_dict: Optional[bool] = None,
931
+ is_padded_inputs: Optional[bool] = True,
932
+ ):
933
+ """If subset_mask is not None, we only want output for the subset of the sequence.
934
+ This means that we only compute the last layer output for these tokens.
935
+ subset_mask: (batch, seqlen), dtype=torch.bool
936
+ """
937
+ hidden_states2 = None
938
+ residual = None
939
+
940
+ for _, layer in enumerate(self.layers):
941
+ if self.gradient_checkpointing and self.training:
942
+
943
+ def create_custom_forward(module):
944
+ def custom_forward(*inputs):
945
+ # None for past_key_value
946
+ return module(*inputs)
947
+
948
+ return custom_forward
949
+
950
+ hidden_states, hidden_states2, residual = torch.utils.checkpoint.checkpoint(
951
+ create_custom_forward(layer),
952
+ hidden_states,
953
+ hidden_states2,
954
+ residual,
955
+ attention_mask,
956
+ None,
957
+ None,
958
+ is_padded_inputs,
959
+ # if you freeze ANY layers, you need `use_reentrant=False`
960
+ # https://github.com/huggingface/transformers/issues/21381
961
+ # https://discuss.pytorch.org/t/checkpoint-with-no-grad-requiring-inputs-problem/19117/7
962
+ use_reentrant=False,
963
+ )
964
+
965
+ else:
966
+ hidden_states, hidden_states2, residual = layer(
967
+ hidden_states,
968
+ hidden_states2,
969
+ residual,
970
+ attention_mask,
971
+ position_ids,
972
+ None,
973
+ is_padded_inputs,
974
+ output_attentions,
975
+ use_cache,
976
+ )
977
+ return hidden_states
978
+
979
+
980
+ class NomicBertPooler(nn.Module):
981
+ def __init__(self, config):
982
+ super().__init__()
983
+ self.dense = nn.Linear(config.n_embd, config.n_embd)
984
+ self.activation = nn.Tanh()
985
+
986
+ def forward(self, hidden_states, pool=True):
987
+ # We "pool" the model by simply taking the hidden state corresponding
988
+ # to the first token.
989
+ first_token_tensor = hidden_states[:, 0] if pool else hidden_states
990
+ pooled_output = self.dense(first_token_tensor)
991
+ pooled_output = self.activation(pooled_output)
992
+ return pooled_output
993
+
994
+
995
+ class NomicBertPredictionHeadTransform(nn.Module):
996
+ def __init__(self, config):
997
+ super().__init__()
998
+ self.dense = nn.Linear(config.n_embd, config.n_embd, bias=config.mlp_fc1_bias)
999
+ approximate = "tanh" if config.activation_function in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"] else "none"
1000
+ if config.activation_function == "swiglu":
1001
+ self.transform_act_fn = F.silu
1002
+ else:
1003
+ self.transform_act_fn = nn.GELU(approximate=approximate)
1004
+
1005
+ self.layer_norm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
1006
+
1007
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
1008
+ hidden_states = self.dense(hidden_states)
1009
+ hidden_states = self.transform_act_fn(hidden_states)
1010
+ hidden_states = self.layer_norm(hidden_states)
1011
+
1012
+ return hidden_states
1013
+
1014
+
1015
+ class NomicBertLMPredictionHead(nn.Module):
1016
+ def __init__(self, config):
1017
+ super().__init__()
1018
+
1019
+ self.transform = NomicBertPredictionHeadTransform(config)
1020
+
1021
+ self.decoder = nn.Linear(config.n_embd, config.vocab_size, bias=config.mlp_fc1_bias)
1022
+
1023
+ def forward(self, hidden_states):
1024
+ hidden_states = self.transform(hidden_states)
1025
+ hidden_states = self.decoder(hidden_states)
1026
+ return hidden_states
1027
+
1028
+
1029
+ class NomicBertPreTrainingHeads(nn.Module):
1030
+ def __init__(self, config):
1031
+ super().__init__()
1032
+ self.predictions = NomicBertLMPredictionHead(config)
1033
+
1034
+ def forward(self, sequence_output):
1035
+ prediction_scores = self.predictions(sequence_output)
1036
+ return prediction_scores
1037
+
1038
+
1039
+ class NomicBertModel(NomicBertPreTrainedModel):
1040
+ def __init__(self, config: GPT2Config, add_pooling_layer=True):
1041
+ super().__init__(config)
1042
+ self.pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
1043
+ if config.vocab_size % self.pad_vocab_size_multiple != 0:
1044
+ config.vocab_size += self.pad_vocab_size_multiple - (config.vocab_size % self.pad_vocab_size_multiple)
1045
+
1046
+ assert config.activation_function in [
1047
+ "gelu",
1048
+ "gelu_new",
1049
+ "gelu_fast",
1050
+ "gelu_pytorch_tanh",
1051
+ "swiglu",
1052
+ "geglu",
1053
+ "glu",
1054
+ ]
1055
+
1056
+ self.embeddings = NomicBertEmbeddings(config)
1057
+ self.emb_drop = nn.Dropout(config.resid_pdrop)
1058
+ self.emb_ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
1059
+ self.encoder = NomicBertEncoder(config)
1060
+ self.pooler = NomicBertPooler(config) if add_pooling_layer else None
1061
+
1062
+ self.apply(partial(_init_weights, initializer_range=config.initializer_range))
1063
+
1064
+ def forward(
1065
+ self,
1066
+ input_ids,
1067
+ attention_mask=None,
1068
+ position_ids=None,
1069
+ token_type_ids=None,
1070
+ return_dict=None,
1071
+ matryoshka_dim=None,
1072
+ ):
1073
+ if token_type_ids is None:
1074
+ token_type_ids = torch.zeros_like(input_ids)
1075
+ hidden_states = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
1076
+ hidden_states = self.emb_ln(hidden_states)
1077
+ hidden_states = self.emb_drop(hidden_states)
1078
+
1079
+ attention_mask = self.get_extended_attention_mask(attention_mask, input_ids.shape)
1080
+ sequence_output = self.encoder(hidden_states, attention_mask=attention_mask, return_dict=return_dict)
1081
+
1082
+ pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
1083
+
1084
+ if matryoshka_dim:
1085
+ sequence_output = sequence_output[:, :matryoshka_dim]
1086
+
1087
+ return BaseModelOutputWithPoolingAndCrossAttentions(
1088
+ last_hidden_state=sequence_output,
1089
+ pooler_output=pooled_output,
1090
+ )
1091
+
1092
+
1093
+ class NomicBertForPreTraining(NomicBertPreTrainedModel):
1094
+ _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
1095
+
1096
+ def __init__(self, config: GPT2Config):
1097
+ super().__init__(config)
1098
+
1099
+ self.bert = NomicBertModel(config, add_pooling_layer=getattr(config, "add_pooling_layer", False))
1100
+ self.cls = NomicBertPreTrainingHeads(config)
1101
+ self.mlm_loss = nn.CrossEntropyLoss()
1102
+
1103
+ # Initialize weights and apply final processing
1104
+ self.apply(partial(_init_weights, initializer_range=config.initializer_range))
1105
+ self.tie_weights()
1106
+
1107
+ def tie_weights(self):
1108
+ self.cls.predictions.decoder.weight = self.bert.embeddings.word_embeddings.weight
1109
+
1110
+ def forward(
1111
+ self,
1112
+ input_ids,
1113
+ position_ids=None,
1114
+ token_type_ids=None,
1115
+ attention_mask=None,
1116
+ labels=None,
1117
+ ):
1118
+ """
1119
+ If labels are provided, they must be -100 for masked out tokens (as specified in the attention
1120
+ mask).
1121
+ Outputs:
1122
+ if `labels` and `next_sentence_label` are not `None`:
1123
+ Outputs the total_loss which is the sum of the masked language modeling loss and the next
1124
+ sentence classification loss.
1125
+ if `labels` or `next_sentence_label` is `None`:
1126
+ Outputs a tuple comprising
1127
+ - the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
1128
+ - the next sentence classification logits of shape [batch_size, 2].
1129
+
1130
+ """
1131
+ outputs = self.bert(
1132
+ input_ids,
1133
+ position_ids=position_ids,
1134
+ token_type_ids=token_type_ids,
1135
+ attention_mask=attention_mask.bool() if attention_mask is not None else None,
1136
+ )
1137
+ sequence_output, _ = outputs.last_hidden_state, outputs.pooler_output
1138
+
1139
+ prediction_scores = self.cls(sequence_output)
1140
+
1141
+ total_loss = None
1142
+ if labels is not None:
1143
+ masked_lm_loss = self.mlm_loss(
1144
+ rearrange(prediction_scores, "... v -> (...) v"),
1145
+ rearrange(labels, "... -> (...)"),
1146
+ )
1147
+ total_loss = masked_lm_loss.float()
1148
+
1149
+ return MaskedLMOutput(
1150
+ loss=total_loss,
1151
+ logits=prediction_scores,
1152
+ hidden_states=outputs.hidden_states,
1153
+ attentions=None,
1154
+ )
1155
+
1156
+
1157
+ class NomicBertForSequenceClassification(NomicBertPreTrainedModel):
1158
+ def __init__(self, config):
1159
+ super().__init__(config)
1160
+ self.num_labels = config.num_labels
1161
+ self.config = config
1162
+
1163
+ self.bert = NomicBertModel(config)
1164
+ classifier_dropout = getattr(config, "classifier_dropout", config.embd_pdrop)
1165
+ self.dropout = nn.Dropout(classifier_dropout)
1166
+ self.classifier = nn.Linear(config.n_embd, config.num_labels)
1167
+
1168
+ # Initialize weights and apply final processing
1169
+ self.post_init()
1170
+
1171
+ def forward(
1172
+ self,
1173
+ input_ids: Optional[torch.Tensor] = None,
1174
+ attention_mask: Optional[torch.Tensor] = None,
1175
+ token_type_ids: Optional[torch.Tensor] = None,
1176
+ position_ids: Optional[torch.Tensor] = None,
1177
+ head_mask: Optional[torch.Tensor] = None,
1178
+ inputs_embeds: Optional[torch.Tensor] = None,
1179
+ labels: Optional[torch.Tensor] = None,
1180
+ output_attentions: Optional[bool] = None,
1181
+ output_hidden_states: Optional[bool] = None,
1182
+ return_dict: Optional[bool] = None,
1183
+ ):
1184
+ r"""
1185
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1186
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1187
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1188
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1189
+ """
1190
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1191
+ outputs = self.bert(
1192
+ input_ids,
1193
+ position_ids=position_ids,
1194
+ token_type_ids=token_type_ids,
1195
+ attention_mask=attention_mask.bool() if attention_mask is not None else None,
1196
+ )
1197
+
1198
+ pooled_output = outputs[1]
1199
+
1200
+ pooled_output = self.dropout(pooled_output)
1201
+ logits = self.classifier(pooled_output)
1202
+
1203
+ loss = None
1204
+ if labels is not None:
1205
+ if self.config.problem_type is None:
1206
+ if self.num_labels == 1:
1207
+ self.config.problem_type = "regression"
1208
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1209
+ self.config.problem_type = "single_label_classification"
1210
+ else:
1211
+ self.config.problem_type = "multi_label_classification"
1212
+
1213
+ if self.config.problem_type == "regression":
1214
+ loss_fct = nn.MSELoss()
1215
+ if self.num_labels == 1:
1216
+ loss = loss_fct(logits.squeeze(), labels.squeeze())
1217
+ else:
1218
+ loss = loss_fct(logits, labels)
1219
+ elif self.config.problem_type == "single_label_classification":
1220
+ loss_fct = nn.CrossEntropyLoss()
1221
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1222
+ elif self.config.problem_type == "multi_label_classification":
1223
+ loss_fct = nn.BCEWithLogitsLoss()
1224
+ loss = loss_fct(logits, labels)
1225
+ if not return_dict:
1226
+ output = (logits,) + outputs[2:]
1227
+ return ((loss,) + output) if loss is not None else output
1228
+
1229
+ return SequenceClassifierOutput(
1230
+ loss=loss,
1231
+ logits=logits,
1232
+ hidden_states=outputs.hidden_states,
1233
+ attentions=outputs.attentions,
1234
+ )
modules.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "idx": 0,
4
+ "name": "0",
5
+ "path": "",
6
+ "type": "sentence_transformers.models.Transformer"
7
+ },
8
+ {
9
+ "idx": 1,
10
+ "name": "1",
11
+ "path": "1_Pooling",
12
+ "type": "sentence_transformers.models.Pooling"
13
+ }
14
+ ]
sentence_bert_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "max_seq_length": 8192,
3
+ "do_lower_case": false
4
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "cls_token": {
3
+ "content": "[CLS]",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "mask_token": {
10
+ "content": "[MASK]",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "[PAD]",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "sep_token": {
24
+ "content": "[SEP]",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ },
30
+ "unk_token": {
31
+ "content": "[UNK]",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false
36
+ }
37
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "added_tokens_decoder": {
3
+ "0": {
4
+ "content": "[PAD]",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false,
9
+ "special": true
10
+ },
11
+ "100": {
12
+ "content": "[UNK]",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false,
17
+ "special": true
18
+ },
19
+ "101": {
20
+ "content": "[CLS]",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false,
25
+ "special": true
26
+ },
27
+ "102": {
28
+ "content": "[SEP]",
29
+ "lstrip": false,
30
+ "normalized": false,
31
+ "rstrip": false,
32
+ "single_word": false,
33
+ "special": true
34
+ },
35
+ "103": {
36
+ "content": "[MASK]",
37
+ "lstrip": false,
38
+ "normalized": false,
39
+ "rstrip": false,
40
+ "single_word": false,
41
+ "special": true
42
+ }
43
+ },
44
+ "clean_up_tokenization_spaces": true,
45
+ "cls_token": "[CLS]",
46
+ "do_lower_case": true,
47
+ "mask_token": "[MASK]",
48
+ "max_length": 8192,
49
+ "model_max_length": 8192,
50
+ "pad_to_multiple_of": null,
51
+ "pad_token": "[PAD]",
52
+ "pad_token_type_id": 0,
53
+ "padding_side": "right",
54
+ "sep_token": "[SEP]",
55
+ "stride": 0,
56
+ "strip_accents": null,
57
+ "tokenize_chinese_chars": true,
58
+ "tokenizer_class": "BertTokenizer",
59
+ "truncation_side": "right",
60
+ "truncation_strategy": "longest_first",
61
+ "unk_token": "[UNK]"
62
+ }
vocab.txt ADDED
The diff for this file is too large to render. See raw diff