jxm commited on
Commit
6ca7132
·
verified ·
1 Parent(s): 166fd0f

Upload ContextualDocumentEmbeddingTransformer

Browse files
Files changed (5) hide show
  1. README.md +199 -0
  2. config.json +28 -0
  3. misc.py +518 -0
  4. model.py +622 -0
  5. model.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,199 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags: []
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+ This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/fsx-checkpoints/jxm/cde/2024-08-06-transductive-pretrain-transductive-long-10node-3/checkpoint-7176",
3
+ "architecture": "transductive",
4
+ "architectures": [
5
+ "ContextualDocumentEmbeddingTransformer"
6
+ ],
7
+ "attn_implementation": null,
8
+ "auto_map": {
9
+ "AutoConfig": "misc.ContextualModelConfig",
10
+ "AutoModel": "model.ContextualDocumentEmbeddingTransformer"
11
+ },
12
+ "cache_dir": null,
13
+ "config_name": null,
14
+ "disable_dropout": true,
15
+ "disable_transductive_rotary_embedding": true,
16
+ "embedder": "nomic-ai/nomic-bert-2048",
17
+ "embedder_rerank": "sentence-transformers/gtr-t5-base",
18
+ "embedding_output_dim": null,
19
+ "limit_layers": null,
20
+ "logit_scale": 50.0,
21
+ "max_seq_length": 512,
22
+ "model_revision": "main",
23
+ "tokenizer_name": null,
24
+ "torch_dtype": "float32",
25
+ "transductive_corpus_size": 512,
26
+ "transductive_sequence_dropout_prob": 0.0,
27
+ "transformers_version": "4.48.0.dev0"
28
+ }
misc.py ADDED
@@ -0,0 +1,518 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Dict, Iterable, List, Optional, Tuple, Union
2
+
3
+ import collections
4
+ import glob
5
+ import json
6
+ import hashlib
7
+ import itertools
8
+ import logging
9
+ import multiprocessing
10
+ import os
11
+ import pickle
12
+ import random
13
+ import requests
14
+ import sys
15
+ import zipfile
16
+
17
+ import datasets
18
+ import numpy as np
19
+ import torch
20
+ import tqdm
21
+ import transformers
22
+
23
+ from cde.lib.dist import get_num_proc, get_rank
24
+
25
+
26
+ def get_cde_cache_dir() -> str:
27
+ script_directory = os.path.normpath(
28
+ os.path.join(
29
+ os.path.dirname(os.path.abspath(__file__)),
30
+ os.pardir, os.pardir,
31
+ )
32
+ )
33
+ return os.path.join(script_directory, "data")
34
+
35
+
36
+ def get_cache_location_from_kwargs(**kwargs):
37
+ cache_location = os.path.join(
38
+ get_cde_cache_dir(), "cluster"
39
+ )
40
+ os.makedirs(cache_location, exist_ok=True)
41
+ return os.path.join(cache_location, md5_hash_kwargs(**kwargs))
42
+
43
+
44
+ def process_qrels_uncached(corpus: datasets.Dataset, qrels: datasets.Dataset) -> Tuple[Dict[str, List[float]], Dict[str, List[str]]]:
45
+ qrels_idxs = collections.defaultdict(list)
46
+ qrels_scores = collections.defaultdict(list)
47
+ corpus_ids = np.array(corpus['_id'])
48
+ skipped_qrels = 0
49
+
50
+ for ex in tqdm.tqdm(qrels, desc='processing qrels', colour='#964B00', leave=False):
51
+ #
52
+ # example:
53
+ # {
54
+ # 'query-id': 1,
55
+ # 'corpus-id': 'b0680508-2019-04-18T13:48:51Z-00002-000',
56
+ # 'score': 2
57
+ # }
58
+ #
59
+ q_id = str(ex['query-id'])
60
+ c_idxs = (corpus_ids == str(ex['corpus-id'])).nonzero()[0]
61
+ #
62
+ assert len(c_idxs) <= 1, f"error - duplicate corpus ID? (found {len(c_idxs)} matches)"
63
+ #
64
+ if len(c_idxs):
65
+ qrels_idxs[q_id].append(c_idxs[0])
66
+ qrels_scores[q_id].append(ex['score'])
67
+ else:
68
+ skipped_qrels += 1
69
+ #
70
+
71
+ if skipped_qrels > 0:
72
+ logging.warning(f'Warning: Skipped {skipped_qrels}/{len(qrels)} qrels.')
73
+
74
+ return qrels_idxs, qrels_scores
75
+
76
+
77
+ def process_qrels(
78
+ corpus: datasets.Dataset, qrels: datasets.Dataset,
79
+ use_cache: bool = True
80
+ ) -> Tuple[Dict[str, List[float]], Dict[str, List[str]]]:
81
+ dataset_cache_file = '_'.join(
82
+ (corpus.cache_files[0]['filename'], qrels.cache_files[0]['filename'])
83
+ )
84
+ cache_file = strip_extension(dataset_cache_file) + '_processed_qrels.p'
85
+ os.makedirs(os.path.dirname(cache_file), exist_ok=True)
86
+
87
+ if not (use_cache and os.path.exists(cache_file)):
88
+ qrels_idxs, qrels_scores = process_qrels_uncached(
89
+ corpus=corpus, qrels=qrels
90
+ )
91
+ if use_cache:
92
+ pickle.dump((qrels_idxs, qrels_scores), open(cache_file, 'wb'))
93
+ else:
94
+ qrels_idxs, qrels_scores = pickle.load(open(cache_file, 'rb'))
95
+
96
+ return qrels_idxs, qrels_scores
97
+
98
+
99
+ def strip_extension(filename: str) -> str:
100
+ """Strips file extension.
101
+
102
+ Ex:
103
+ >> strip_extension('/root/dir/sub/file.ext')
104
+ '/root/dir/sub/file'
105
+ """
106
+ return os.path.splitext(filename)[0]
107
+
108
+
109
+ def md5_hash(t: Tuple[str]) -> str:
110
+ return hashlib.md5('__'.join(t).encode()).hexdigest()
111
+
112
+
113
+ def md5_hash_kwargs(**kwargs) -> str:
114
+ # We ignore special hf args that start with _ like '__cached__setup_devices'.
115
+ safe_kwargs = {k: str(v) for k,v in kwargs.items() if not k.startswith('_')}
116
+ s = json.dumps(safe_kwargs, sort_keys=True)
117
+ return hashlib.md5(s.encode()).hexdigest()
118
+
119
+ def download_url(url: str, save_path: str, chunk_size: int = 1024):
120
+ """Download url with progress bar using tqdm
121
+ https://stackoverflow.com/questions/15644964/python-progress-bar-and-downloads
122
+ Args:
123
+ url (str): downloadable url
124
+ save_path (str): local path to save the downloaded file
125
+ chunk_size (int, optional): chunking of files. Defaults to 1024.
126
+ """
127
+ r = requests.get(url, stream=True)
128
+ total = int(r.headers.get('Content-Length', 0))
129
+ with open(save_path, 'wb') as fd, tqdm.tqdm(
130
+ desc=save_path,
131
+ total=total,
132
+ unit='iB',
133
+ unit_scale=True,
134
+ unit_divisor=chunk_size,
135
+ ) as bar:
136
+ for data in r.iter_content(chunk_size=chunk_size):
137
+ size = fd.write(data)
138
+ bar.update(size)
139
+
140
+
141
+ def unzip(zip_file: str, out_dir: str):
142
+ print("unzipping =>", zip_file)
143
+ zip_ = zipfile.ZipFile(zip_file, "r")
144
+ zip_.extractall(path=out_dir)
145
+ zip_.close()
146
+
147
+
148
+ def download_url_and_unzip(url: str, out_dir: str, chunk_size: int = 1024) -> str:
149
+ os.makedirs(out_dir, exist_ok=True)
150
+ dataset = url.split("/")[-1]
151
+ zip_file = os.path.join(out_dir, dataset)
152
+
153
+ if not os.path.isfile(zip_file):
154
+ logging.info("Downloading {} ...".format(dataset))
155
+ download_url(url, zip_file, chunk_size)
156
+
157
+ if not os.path.isdir(zip_file.replace(".zip", "")):
158
+ logging.info("Unzipping {} ...".format(dataset))
159
+ unzip(zip_file, out_dir)
160
+
161
+ return os.path.join(out_dir, dataset.replace(".zip", ""))
162
+
163
+
164
+ def tqdm_if_main_worker(iterable: Iterable, **kwargs) -> Iterable:
165
+ if get_rank() == 0:
166
+ return tqdm.tqdm(iterable, **kwargs)
167
+ else:
168
+ return iterable
169
+
170
+
171
+ class ContextualModelConfig(transformers.configuration_utils.PretrainedConfig):
172
+ """We create a dummy configuration class that will just set properties
173
+ based on whatever kwargs we pass in.
174
+
175
+ When this class is initialized (see experiments.py) we pass in the
176
+ union of all data, model, and training args, all of which should
177
+ get saved to the config json.
178
+ """
179
+
180
+ def __init__(self, **kwargs):
181
+ for key, value in kwargs.items():
182
+ try:
183
+ json.dumps(value)
184
+ setattr(self, key, value)
185
+ except TypeError:
186
+ # value was not JSON-serializable, skip
187
+ continue
188
+ super().__init__()
189
+
190
+
191
+ def independent_crop(
192
+ input_ids: torch.Tensor, pad_token_id: int,
193
+ l1: int = 256, l2: int = 256) -> Tuple[torch.Tensor, torch.Tensor]:
194
+ """Returns two independent crops from input_ids.
195
+
196
+ Assumes input_ids has a beginning and end token, like
197
+ [101, ..., 102, 0, 0, 0].
198
+
199
+ Args:
200
+ input_ids: tensor of IDs
201
+ pad_token_id: ID of pad tokens in input_ids
202
+ l1: length of span 1, cropped
203
+ l2: length of span 2, cropped
204
+ Returns:
205
+ span1: first crop (of length l1)
206
+ span2: second crop (of length l2)
207
+ """
208
+ # Count tokens until pad.
209
+ if (input_ids == pad_token_id).sum() == 0:
210
+ N = len(input_ids)
211
+ else:
212
+ N = (input_ids == pad_token_id).int().argmax().item()
213
+
214
+ ####
215
+ ###
216
+ ##
217
+ ## Contriever: We use the random cropping data
218
+ ## augmentation, with documents of 256 tokens and span
219
+ ## sizes sampled between 5% and 50% of the document
220
+ ## length
221
+ ##
222
+ ###
223
+ #####
224
+ ####### LaPraDor: The maximum lengths set for queries and
225
+ ####### documents are 64 and 350...
226
+ #####
227
+ # TODO is this divide-by-two a good idea? (Don't want s1=s2 ever..)
228
+ nl1 = min(N//2, l1)
229
+ nl2 = min(N//2, l2)
230
+
231
+ s1_start = random.randint(1, N-nl1)
232
+ s2_start = random.randint(1, N-nl2)
233
+
234
+ s1_idxs = itertools.chain(
235
+ [0], range(s1_start, s1_start+nl1), [N-1]
236
+ )
237
+ s1 = input_ids[torch.tensor(list(s1_idxs))]
238
+ s2_idxs = itertools.chain(
239
+ [0], range(s2_start, s2_start+nl2), [N-1]
240
+ )
241
+ s2 = input_ids[torch.tensor(list(s2_idxs))]
242
+ return (s1, s2)
243
+
244
+
245
+ def load_dataset_tables(
246
+ files: Iterable[str], num_workers: int = 16
247
+ ) -> Iterable[datasets.table.MemoryMappedTable]:
248
+ import concurrent
249
+ from multiprocessing import Pool
250
+
251
+ # num_workers = min(num_workers, len(files))
252
+ num_workers = min(32, len(files))
253
+
254
+ use_threads = True
255
+ if use_threads:
256
+ pool_cls = concurrent.futures.ThreadPoolExecutor
257
+ pool_kwargs = {"max_workers": num_workers}
258
+ else:
259
+ pool_cls = Pool
260
+ pool_kwargs = {"processes": num_workers}
261
+
262
+ with pool_cls(**pool_kwargs) as pool:
263
+ if len(files) > 10:
264
+ files = tqdm_if_main_worker(
265
+ files,
266
+ desc=f"Loading {len(files)} files with {num_workers} workers",
267
+ total=len(files),
268
+ colour="#ffbd88"
269
+ )
270
+
271
+ result = list(
272
+ pool.map(datasets.table.MemoryMappedTable.from_file, files)
273
+ )
274
+ return result
275
+
276
+
277
+ def datasets_fast_load_from_disk(cache_path: str) -> datasets.Dataset:
278
+ logging.info(f"fast_load_from_disk called with path:", cache_path)
279
+ dataset_info_path = os.path.join(cache_path, "dataset_info.json")
280
+ with open(dataset_info_path, encoding="utf-8") as dataset_info_file:
281
+ dataset_info = datasets.DatasetInfo.from_dict(json.load(dataset_info_file))
282
+
283
+ dataset_state_path = os.path.join(cache_path, "state.json")
284
+ with open(dataset_state_path, encoding="utf-8") as state_file:
285
+ state = json.load(state_file)
286
+
287
+ files = glob.glob(os.path.join(cache_path, "data-*.arrow"))
288
+ files = sorted(files)
289
+ num_workers = get_num_proc()
290
+ ds_tables = load_dataset_tables(
291
+ files=files,
292
+ num_workers=num_workers
293
+ )
294
+ arrow_table = datasets.table.concat_tables(ds_tables)
295
+
296
+ split = state["_split"]
297
+ split = datasets.splits.Split(split) if split is not None else split
298
+
299
+ # print("returning dataset")
300
+ return datasets.Dataset(
301
+ arrow_table=arrow_table,
302
+ info=dataset_info,
303
+ split=split,
304
+ fingerprint=state["_fingerprint"],
305
+ )
306
+
307
+
308
+ def tokenize_dataset(
309
+ dataset: datasets.Dataset,
310
+ tokenizer: transformers.PreTrainedTokenizer,
311
+ max_length: int,
312
+ text_key: str,
313
+ padding_strategy: str
314
+ ) -> datasets.Dataset:
315
+ def tokenize_text(ex: Dict) -> Dict:
316
+ tt = tokenizer(
317
+ ex[text_key],
318
+ max_length=max_length,
319
+ truncation=True,
320
+ padding=padding_strategy,
321
+ )
322
+ for k,v in tt.items():
323
+ ex[f"{text_key}_{k}"] = v
324
+ ex["length"] = [len(tt) for tt in ex[f"{text_key}_input_ids"]]
325
+ return ex
326
+
327
+ # generate unique hash for tokenizer
328
+ vocab = tokenizer.vocab
329
+ vocab_words = tuple(sorted(vocab.keys(), key=lambda word: vocab[word]))
330
+ vocab_hash = md5_hash(vocab_words)
331
+
332
+ data_fingerprint = '__'.join((
333
+ dataset._fingerprint, str(vocab_hash), str(max_length),
334
+ text_key, padding_strategy
335
+ ))
336
+ data_fingerprint = md5_hash(data_fingerprint)
337
+ dataset = dataset.map(
338
+ tokenize_text,
339
+ new_fingerprint=data_fingerprint,
340
+ batched=True,
341
+ load_from_cache_file=True,
342
+ )
343
+ return dataset
344
+
345
+
346
+ class TensorRunningAverages:
347
+ _store_sum: Dict[str, torch.Tensor]
348
+ _store_total: Dict[str, torch.Tensor]
349
+
350
+ def __init__(self):
351
+ self._store_sum = {}
352
+ self._store_total = {}
353
+
354
+ def __iter__(self) -> Iterable[str]:
355
+ return iter(self._store_sum.keys())
356
+
357
+ def update(self, key: str, val: Union[int, float, torch.Tensor]) -> None:
358
+ if key not in self._store_sum:
359
+ self.clear(key)
360
+ if isinstance(val, torch.Tensor):
361
+ val = val.item() # tensor -> num
362
+ self._store_sum[key] += val
363
+ self._store_total[key] += 1
364
+
365
+ def get(self, key: str) -> float:
366
+ total = max(self._store_total.get(key).item(), 1.0)
367
+ return (self._store_sum[key] / float(total)).item() or 0.0
368
+
369
+ def clear(self, key: str) -> None:
370
+ self._store_sum[key] = torch.tensor(0.0, dtype=torch.float32)
371
+ self._store_total[key] = torch.tensor(0, dtype=torch.int32)
372
+
373
+ def clear_all(self) -> None:
374
+ for key in self._store_sum:
375
+ self.clear(key)
376
+
377
+ def get_and_clear_all(self) -> Dict[str, float]:
378
+ metrics = {}
379
+ for key in self:
380
+ metrics[key] = self.get(key)
381
+ self.clear(key)
382
+ return metrics
383
+
384
+ def load_embedder_and_tokenizer(name: str) -> Tuple[
385
+ transformers.PreTrainedModel,
386
+ transformers.PreTrainedTokenizer
387
+ ]:
388
+ if name.startswith("nomic") or (name == "bert-base-uncased"):
389
+ from cde.lib.nomic_bert import NomicBertModel
390
+ if name.endswith("--from-scratch"):
391
+ name = name.replace("--from-scratch", "")
392
+ config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
393
+ model = NomicBertModel._from_config(config)
394
+ else:
395
+ model = NomicBertModel.from_pretrained(
396
+ name, add_pooling_layer=False
397
+ )
398
+ tokenizer = transformers.AutoTokenizer.from_pretrained(name)
399
+ elif name in ["gtr-base", "gtr_base"]:
400
+ model = transformers.AutoModel.from_pretrained(
401
+ "sentence-transformers/gtr-t5-base"
402
+ ).encoder
403
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
404
+ "sentence-transformers/gtr-t5-base"
405
+ )
406
+ elif name == "pile-t5-base-encoder":
407
+ model = transformers.AutoModel.from_pretrained(
408
+ "EleutherAI/pile-t5-base"
409
+ ).encoder
410
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
411
+ "EleutherAI/pile-t5-base"
412
+ )
413
+ tokenizer.pad_token = tokenizer.eos_token
414
+ elif name == "pile-t5-base-decoder":
415
+ model = transformers.AutoModel.from_pretrained(
416
+ "EleutherAI/pile-t5-base"
417
+ ).decoder
418
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
419
+ "EleutherAI/pile-t5-base"
420
+ )
421
+ tokenizer.pad_token = tokenizer.eos_token
422
+ elif name.startswith("gpt2") or name.startswith("meta-llama") or ("Llama" in name):
423
+ model = transformers.AutoModelForCausalLM.from_pretrained(
424
+ name,
425
+ # torch_dtype=torch.bfloat16,
426
+ attn_implementation="flash_attention_2" if torch.cuda.is_available() else "sdpa",
427
+ low_cpu_mem_usage=True,
428
+ # device_map="auto",
429
+ )
430
+ model.padding_side = "right"
431
+ tokenizer = transformers.AutoTokenizer.from_pretrained(name)
432
+ tokenizer.pad_token = tokenizer.eos_token
433
+ tokenizer.add_eos_token = True
434
+ elif "Modern" in name:
435
+ print("special loading for ModernBERT!")
436
+ # [1] needed for faster training
437
+ # model = transformers.AutoModel.from_pretrained(name, trust_remote_code=True, reference_compile=True)
438
+ # [2] needed for non-breaking inference
439
+ model = transformers.AutoModel.from_pretrained(name, trust_remote_code=True, reference_compile=False)
440
+ tokenizer = transformers.AutoTokenizer.from_pretrained(name)
441
+ else:
442
+ model = transformers.AutoModel.from_pretrained(name, trust_remote_code=True)
443
+ tokenizer = transformers.AutoTokenizer.from_pretrained(name)
444
+ return model, tokenizer
445
+
446
+
447
+ def inputs_for_key(inputs: Dict[str, torch.Tensor], key: str):
448
+ key += "_"
449
+ return {k.replace(key, ""): v for k,v in inputs.items() if k.startswith(key)}
450
+
451
+
452
+ def count_cpus() -> int:
453
+ try:
454
+ return len(os.sched_getaffinity(0))
455
+ except AttributeError:
456
+ return multiprocessing.cpu_count()
457
+
458
+
459
+ def shuffle_batches(g: torch.Generator, list_of_tensors: List[torch.Tensor]) -> List[int]:
460
+ all_indices = []
461
+ for batch_tensor in tqdm_if_main_worker(list_of_tensors, colour="green", desc="Sampler shuffling per-batch"):
462
+ rand_perm = torch.randperm(len(batch_tensor), generator=g)
463
+ batch_list = batch_tensor[rand_perm].tolist()
464
+ all_indices.extend(batch_list)
465
+ return all_indices
466
+
467
+
468
+ # def shuffle_batches_multiproc(g: torch.Generator, list_of_tensors: List[torch.Tensor], num_processes: int = 8) -> List[int]:
469
+ # all_indices = []
470
+ # print(f"Shuffling {len(list_of_tensors)} tensors with {num_processes} workers.")
471
+ # pbar = tqdm_if_main_worker(list_of_tensors, colour="orange", desc=f"Sampler shuffling per-batch (nproc={num_processes})")
472
+ # pool = multiprocessing.Pool(processes=num_processes)
473
+ # chunk_size = len(list_of_tensors) // num_processes
474
+ # chunks = [list_of_tensors[i:i + chunk_size] for i in range(0, len(list_of_tensors), chunk_size)]
475
+ # worker_func = functools.partial(shuffle_batches, g=g)
476
+ # results = pool.map(worker_func, chunks)
477
+ # all_indices = []
478
+ # for result in results:
479
+ # all_indices.extend(result)
480
+ # pbar.update()
481
+ # return all_indices
482
+
483
+
484
+ def exit_if_running_or_finished_wandb(
485
+ project_name: str,
486
+ exp_group: str, exp_name: str
487
+ ) -> None:
488
+ print("Checking if experiment is already running...")
489
+ import wandb
490
+
491
+ api = wandb.Api()
492
+ running_runs = api.runs(
493
+ path="cde-0",
494
+ filters={
495
+ "display_name": exp_name,
496
+ "state": {"$regex": "Running|Finished"},
497
+ "config.exp_group": exp_group,
498
+ }
499
+ )
500
+ print("Found", len(running_runs), f"runs with name {exp_name} and group {exp_group} in {project_name}.")
501
+
502
+ if len(running_runs) > 0:
503
+ print("Exiting because experiment is already running or completed.")
504
+ sys.exit(0)
505
+
506
+
507
+ HN_FILTER_TOKENIZER_MAP = {
508
+ "nomic": "nomic-ai/nomic-embed-text-v1",
509
+ "stella": "dunzhang/stella_en_400M_v5",
510
+ "sbert": "sentence-transformers/all-MiniLM-L6-v2",
511
+ "sentence_t5": "sentence-transformers/sentence-t5-base",
512
+ "gte": "Alibaba-NLP/gte-large-en-v1.5",
513
+ }
514
+ def load_hn_filter_tokenizer(tokenizer_name: str) -> Optional[transformers.PreTrainedTokenizer]:
515
+ if tokenizer_name in HN_FILTER_TOKENIZER_MAP:
516
+ return transformers.AutoTokenizer.from_pretrained(HN_FILTER_TOKENIZER_MAP[tokenizer_name])
517
+ else:
518
+ return None
model.py ADDED
@@ -0,0 +1,622 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional
2
+
3
+ import copy
4
+ import torch
5
+ import torch.nn as nn
6
+ import transformers
7
+
8
+ from cde.lib.dist import print0
9
+ from cde.lib.tensor import mean_pool, mean_pool_3d, mean_pool_weighted, last_token_pool
10
+
11
+ from cde.lib import load_embedder_and_tokenizer, ContextualModelConfig
12
+
13
+
14
+ gpt_tokenizer = transformers.AutoTokenizer.from_pretrained("gpt2")
15
+
16
+ def limit_layers(model: transformers.PreTrainedModel, n_layers: int) -> None:
17
+ if hasattr(model, 'transformer'):
18
+ if hasattr(model.transformer, 'h'):
19
+ # gpt2
20
+ model.transformer.h = model.transformer.h[:n_layers]
21
+ else:
22
+ model.transformer.layer = model.transformer.layer[:n_layers]
23
+ elif hasattr(model, 'encoder'):
24
+ if hasattr(model.encoder, 'layers'):
25
+ model.encoder.layers = model.encoder.layers[:n_layers]
26
+ else:
27
+ model.encoder.layer = model.encoder.layer[:n_layers]
28
+ else:
29
+ raise RuntimeError(f"unknown how to limit layers of model {type(model)}")
30
+
31
+
32
+ def disable_dropout(model: torch.nn.Module):
33
+ dropout_modules = [m for m in model.modules() if isinstance(m, torch.nn.Dropout)]
34
+ for m in dropout_modules:
35
+ m.p = 0.0
36
+ print0(
37
+ f"Disabled {len(dropout_modules)} dropout modules from model type {type(model)}"
38
+ )
39
+
40
+
41
+ def disable_causality(model: torch.nn.Module):
42
+ disabled_modules = 0
43
+ for m in model.modules():
44
+ if hasattr(m, "is_causal"):
45
+ m.is_causal = False
46
+ disabled_modules += 1
47
+ print0(
48
+ f"Set is_causal=False in {disabled_modules} modules from model type {type(model)}"
49
+ )
50
+
51
+
52
+ class ContextualModelMixin(nn.Module):
53
+ @property
54
+ def num_corpus_tokens(self) -> int:
55
+ return self.transductive_corpus_size * self.transductive_tokens_per_document
56
+
57
+ def contextual_init(self):
58
+ self.n_soft_prompt = 8
59
+ self.prompt_projection = torch.nn.Sequential(
60
+ torch.nn.Linear(self.hidden_size, self.hidden_size),
61
+ torch.nn.ReLU(),
62
+ torch.nn.Linear(self.hidden_size, self.hidden_size * self.n_soft_prompt)
63
+ )
64
+ self.transductive_corpus_size = vars(self.config).get("transductive_corpus_size", 1)
65
+ self.transductive_tokens_per_document = vars(self.config).get("transductive_tokens_per_document", 1)
66
+ self.randomize_dataset_sequence_order = True
67
+ self.sequence_dropout_prob = vars(self.config).get("transductive_sequence_dropout_prob", 0.0)
68
+ if self.sequence_dropout_prob > 0.0:
69
+ self.sequence_dropout_null_embedding = torch.nn.Parameter(
70
+ torch.randn(self.hidden_size) * 0.01,
71
+ requires_grad = True
72
+ )
73
+ self.output_projection = torch.nn.Sequential(
74
+ torch.nn.Linear(self.hidden_size, self.hidden_size),
75
+ torch.nn.ReLU(),
76
+ torch.nn.Linear(self.hidden_size, self.hidden_size)
77
+ )
78
+
79
+ def _prepare_dataset_embeddings(
80
+ self,
81
+ input_ids: torch.Tensor,
82
+ dataset_embeddings: torch.Tensor,
83
+ null_dataset_embedding: bool = False,
84
+ ) -> torch.Tensor:
85
+ if not isinstance(dataset_embeddings, torch.Tensor):
86
+ dataset_embeddings = torch.tensor(dataset_embeddings)
87
+
88
+ if len(dataset_embeddings.shape) == 2:
89
+ # Auto-expand for a batch.
90
+ dataset_embeddings = dataset_embeddings[None, :, :] # (b, d) -> (1, b, d)
91
+ dataset_embeddings = dataset_embeddings.to(input_ids.device)
92
+
93
+ batch_size = input_ids.shape[0]
94
+ if (self.transductive_tokens_per_document > 1):
95
+ if self.training:
96
+ # Choose N random documents to fill our context window with.
97
+ # This logic is a little confusing but allows us to sample a
98
+ # different batch *per-document*
99
+ assert dataset_embeddings.shape[1] == self.transductive_tokens_per_document
100
+ R = torch.randint(
101
+ low=0,
102
+ high=len(dataset_embeddings),
103
+ size=(batch_size, self.config.transductive_corpus_size),
104
+ device=dataset_embeddings.device
105
+ )
106
+ # TODO make this deterministic somehow for evaluation?
107
+ dataset_embeddings = dataset_embeddings[R].reshape((batch_size, self.num_corpus_tokens, self.hidden_size))
108
+ else:
109
+ dataset_embeddings = dataset_embeddings.reshape((1, self.num_corpus_tokens, self.hidden_size))
110
+ # print("reshaped to dataset_embeddings.shape =", dataset_embeddings.shape)
111
+
112
+ if dataset_embeddings.shape[1] > self.num_corpus_tokens:
113
+ # If too many dataset embeddings are passed in, just take the first N until
114
+ # we have the proper number.
115
+ dataset_embeddings = dataset_embeddings[:, :self.num_corpus_tokens, :]
116
+
117
+ _, corpus_size, _hidden_size = dataset_embeddings.shape
118
+ if _ == 1:
119
+ # Auto-expand for a batch.
120
+ dataset_embeddings = dataset_embeddings.expand((batch_size, -1, -1))
121
+
122
+ if self.training and self.sequence_dropout_prob > 0.0:
123
+ sequence_dropout_mask = (
124
+ torch.rand((batch_size, corpus_size), device=dataset_embeddings.device) < self.sequence_dropout_prob
125
+ )
126
+ null_embeddings = self.sequence_dropout_null_embedding[None, None].expand(batch_size, corpus_size, -1)
127
+ dataset_embeddings = torch.where(
128
+ sequence_dropout_mask[..., None], null_embeddings, dataset_embeddings
129
+ )
130
+ elif null_dataset_embedding:
131
+ null_embeddings = self.sequence_dropout_null_embedding[None, None].expand(batch_size, corpus_size, -1)
132
+ dataset_embeddings = null_embeddings
133
+
134
+ # backbone_max_seq_length = self.backbone.config.max_trained_positions
135
+ # assert batch_size + (2 * self.n_soft_prompt + corpus_size) <= backbone_max_seq_length, "too many hard negatives for backbone model"
136
+ soft_prompt = torch.ones((1, self.hidden_size), device=dataset_embeddings.device, dtype=dataset_embeddings.dtype)
137
+ soft_prompt = self.prompt_projection(soft_prompt).reshape((1, self.n_soft_prompt, self.hidden_size))
138
+ soft_prompt = soft_prompt.expand((len(dataset_embeddings), -1, -1)) # -> (b, 4+b, d) # soft_prompt.repeat((len(input_ids), 1, 1))
139
+ soft_prompt = torch.cat((dataset_embeddings, soft_prompt), dim=1)
140
+
141
+ return soft_prompt
142
+
143
+
144
+ class BiEncoder(transformers.PreTrainedModel):
145
+ config_class = ContextualModelConfig
146
+ embedder: transformers.PreTrainedModel
147
+ def __init__(
148
+ self,
149
+ config, #: transformers.PreTrainedConfig,
150
+ ):
151
+ super().__init__(config=config)
152
+ embedder, _ = load_embedder_and_tokenizer(
153
+ config.embedder,
154
+ )
155
+
156
+ if config.limit_layers:
157
+ print0(f"Limiting layers to {config.limit_layers}")
158
+ limit_layers(embedder, config.limit_layers)
159
+
160
+ self.embedder = embedder
161
+ # if ("t5" in embedder.config.model_type):
162
+ # print0(f"using torch.compile() on embedder of type `{embedder.config.model_type}`")
163
+ # self.embedder = torch.compile(self.embedder)
164
+ self.hidden_size = self.embedder.config.hidden_size
165
+ # Allow pooling to multiple tokens per document
166
+ self.transductive_tokens_per_document = vars(self.config).get("transductive_tokens_per_document", 1)
167
+ self.mlp = torch.nn.Sequential(
168
+ torch.nn.Linear(self.hidden_size, self.hidden_size),
169
+ torch.nn.GELU(),
170
+ torch.nn.Linear(self.hidden_size, self.config.embedding_output_dim or self.hidden_size),
171
+ )
172
+ self.temp = config.logit_scale
173
+
174
+ if config.disable_dropout:
175
+ disable_dropout(self)
176
+ self.pooling_strategy = vars(config).get("pooling_strategy", "mean")
177
+
178
+ def forward(
179
+ self,
180
+ input_ids: torch.Tensor,
181
+ attention_mask: torch.Tensor,
182
+ dataset_input_ids: Optional[torch.Tensor] = None,
183
+ dataset_attention_mask: Optional[torch.Tensor] = None,
184
+ token_type_ids = None,
185
+ output_hidden_states: bool = False,
186
+ ) -> torch.Tensor:
187
+ """
188
+ query_embedding (float torch.Tensor) - shape (batch_size, embedding_dim)
189
+ document_embeddings (float torch.Tensor) - shape (corpus_size, embedding_dim)
190
+ where the corpus_size >= batch_size and is structured like this:
191
+ [d1, d2, d3, hn1_1, hn1_2, hn2_1, hn2_2, hn3_1, hn3_2]
192
+ for a corpus with three documents and two hard negatives per document
193
+ """
194
+ del token_type_ids
195
+
196
+ outputs = (
197
+ self.embedder(
198
+ input_ids=input_ids,
199
+ attention_mask=attention_mask,
200
+ ).last_hidden_state
201
+ )
202
+ if self.transductive_tokens_per_document > 1:
203
+ document_embeddings = None
204
+ batch_size, seq_length, output_dim = outputs.shape
205
+
206
+ if seq_length % self.transductive_tokens_per_document != 0:
207
+ # Pad to nearest multiple
208
+ n_extra_embeds = self.transductive_tokens_per_document - (seq_length % self.transductive_tokens_per_document)
209
+ outputs = torch.cat(
210
+ (outputs, torch.zeros((batch_size, n_extra_embeds, output_dim), device=outputs.device)),
211
+ dim=1
212
+ )
213
+ attention_mask = torch.cat(
214
+ (attention_mask, torch.zeros((batch_size, n_extra_embeds), device=attention_mask.device)),
215
+ dim=1
216
+ )
217
+ seq_length += n_extra_embeds
218
+ print(f"Added {n_extra_embeds} padding tokens to input_ids and attention_mask")
219
+
220
+ # print("ftransductive_tokens_per_document {self.transductive_tokens_per_document} outputs.shape =", outputs.shape)
221
+
222
+ outputs = outputs.reshape(
223
+ (batch_size, self.transductive_tokens_per_document, seq_length // self.transductive_tokens_per_document, output_dim)
224
+ )
225
+
226
+ attention_mask = attention_mask.reshape((batch_size, self.transductive_tokens_per_document, -1))
227
+ document_embeddings = mean_pool_3d(outputs, attention_mask)
228
+
229
+ document_embeddings = document_embeddings.reshape((batch_size, self.transductive_tokens_per_document, output_dim))
230
+ else:
231
+ if self.pooling_strategy == "mean":
232
+ document_embeddings = mean_pool(outputs, attention_mask)
233
+ else:
234
+ document_embeddings = document_embeddings.max(dim=1)
235
+ output = self.mlp(document_embeddings)
236
+ # breakpoint()
237
+
238
+ if output_hidden_states:
239
+ return {
240
+ "hidden_states": outputs,
241
+ "pooled": output,
242
+ }
243
+ else:
244
+ return output
245
+
246
+
247
+ class DatasetConditionedAutoregressive(transformers.PreTrainedModel, ContextualModelMixin):
248
+ def __init__(
249
+ self,
250
+ config,
251
+ dataset_backbone: transformers.PreTrainedModel,
252
+ first_stage_hidden_size: int,
253
+ ):
254
+ super().__init__(config=config)
255
+ self.backbone = dataset_backbone
256
+ self.backbone_hidden_size = self.backbone.config.hidden_size
257
+ self.hidden_size = first_stage_hidden_size # Input token size
258
+ self.contextual_init()
259
+ disable_causality(self.backbone)
260
+
261
+ self.pool_ignore_contextual_tokens = vars(self.config).get("pool_ignore_contextual_tokens", False)
262
+ self.pool_ignore_instruction_tokens = vars(self.config).get("pool_ignore_instruction_tokens", False)
263
+ self.pool_instruction_end_id = self.backbone.config.bos_token_id
264
+
265
+ # Override contextual init
266
+ self.output_projection = torch.nn.Sequential(
267
+ torch.nn.Linear(self.backbone_hidden_size, self.backbone_hidden_size),
268
+ torch.nn.ReLU(),
269
+ torch.nn.Linear(self.backbone_hidden_size, self.backbone_hidden_size)
270
+ )
271
+ self._shift_rotary_embedding()
272
+
273
+ @property
274
+ def num_corpus_tokens(self) -> int:
275
+ return self.config.transductive_corpus_size * self.transductive_tokens_per_document
276
+
277
+ @property
278
+ def corpus_token_ratio(self) -> float:
279
+ # How many tokens from the first stage make one token in the second
280
+ # stage?
281
+ return self.backbone_hidden_size / self.hidden_size
282
+
283
+ def corpus_token_pad_size(self, n_tokens: int) -> int:
284
+ return self.hidden_size % self.backbone_hidden_size
285
+
286
+ def _shift_rotary_embedding(self) -> None:
287
+ disable_transductive_rotary_embedding = vars(self.config).get("disable_transductive_rotary_embedding", True)
288
+ # TODO: Can we do this for LLAMA?
289
+ print0("Warning: Positional embedding disabling not implemented for LLAMA.")
290
+
291
+ def forward(
292
+ self,
293
+ input_ids: torch.Tensor,
294
+ attention_mask: torch.Tensor,
295
+ dataset_embeddings: torch.Tensor,
296
+ output_hidden_states: bool = False,
297
+ null_dataset_embedding: bool = False,
298
+ ) -> torch.Tensor:
299
+ soft_prompt = self._prepare_dataset_embeddings(
300
+ input_ids=input_ids,
301
+ dataset_embeddings=dataset_embeddings,
302
+ null_dataset_embedding=null_dataset_embedding,
303
+ )
304
+
305
+ # Reshape for this model.
306
+ # print("[DatasetConditionedAutoregressive] 1 -> soft_prompt.shape =", soft_prompt.shape)
307
+ num_soft_elements = torch.prod(torch.tensor(soft_prompt.shape[1:])).item()
308
+ soft_prompt = soft_prompt.reshape((soft_prompt.shape[0], num_soft_elements))
309
+ num_padding_elements = self.backbone_hidden_size - (num_soft_elements % self.backbone_hidden_size)
310
+ padding = torch.ones((soft_prompt.shape[0], num_padding_elements), device=soft_prompt.device)
311
+ soft_prompt = torch.cat((soft_prompt, padding), dim=1)
312
+ soft_prompt = soft_prompt.reshape(
313
+ (soft_prompt.shape[0], -1, self.backbone_hidden_size)
314
+ )
315
+ # print("[DatasetConditionedAutoregressive] 2 -> soft_prompt.shape =", soft_prompt.shape)
316
+
317
+ backbone_attention_mask = torch.ones(
318
+ soft_prompt.shape[0:2],
319
+ dtype=torch.long,
320
+ device=soft_prompt.device,
321
+ )
322
+ token_embeddings = self.backbone.get_input_embeddings()
323
+ inputs_embeds = token_embeddings(input_ids) # (b, s) -> (b, s, d)
324
+ # print("[2] inputs_embeds.shape =", inputs_embeds.shape)
325
+ inputs_embeds = torch.cat((soft_prompt, inputs_embeds), dim=1) # (v, 4+b+s, d)
326
+ # print("[3.a] inputs_embeds.shape =", inputs_embeds.shape)
327
+ input_attention_mask = torch.cat((backbone_attention_mask, attention_mask), dim=1)
328
+ # print("[3.b] attention_mask.shape =", attention_mask.shape)
329
+
330
+ output = self.backbone(
331
+ inputs_embeds=inputs_embeds,
332
+ attention_mask=input_attention_mask,
333
+ output_hidden_states=True,
334
+ ) # (1, 4 + b + s, d)
335
+ # trim soft prompt
336
+ output_vectors = output.hidden_states[-1]
337
+ n_soft_prompt_tokens = soft_prompt.shape[1]
338
+
339
+ if self.pool_ignore_instruction_tokens:
340
+ # Denote the end of an instruction with an extra BOS token.
341
+ # This is a bit arcane but relies on the fact that there will be a BOS token after the
342
+ # instruction, but also there may or may not be a BOS token at the beginning.
343
+ instruction_end_idx = (
344
+ (input_ids == self.pool_instruction_end_id) &
345
+ attention_mask &
346
+ (torch.arange(input_ids.shape[1], device=input_ids.device)[None, :] > 0)
347
+ ).int().argmax(1)
348
+ is_instruction_token_mask = (
349
+ torch.arange(input_ids.shape[1], device=input_ids.device)[None, :] <= instruction_end_idx[:, None]
350
+ )
351
+ # catch edge case where there is no instruction
352
+ is_instruction_token_mask = is_instruction_token_mask.where(
353
+ (instruction_end_idx > 0)[:, None], torch.zeros_like(is_instruction_token_mask)
354
+ )
355
+ input_attention_mask = torch.cat((
356
+ backbone_attention_mask,
357
+ attention_mask & ~is_instruction_token_mask), dim=1
358
+ )
359
+
360
+ output_attention_mask = input_attention_mask
361
+ if self.pool_ignore_contextual_tokens:
362
+ output_vectors = output_vectors[:, n_soft_prompt_tokens:, :]
363
+ output_attention_mask = output_attention_mask[:, n_soft_prompt_tokens:]
364
+
365
+ # Take last token position
366
+ if vars(self.config).get("pooling_strategy") == "last_token":
367
+ output_pooled = last_token_pool(output_vectors, output_attention_mask)
368
+ elif vars(self.config).get("pooling_strategy") == "mean":
369
+ output_pooled = mean_pool(output_vectors, output_attention_mask)
370
+ else:
371
+ output_pooled = mean_pool_weighted(output_vectors, output_attention_mask)
372
+
373
+ # average with original vectors
374
+ output = self.output_projection(output_pooled) # (b, 2d) -> (b, d)
375
+
376
+ if output_hidden_states:
377
+ return {
378
+ "hidden_states": output_vectors,
379
+ "pooled": output,
380
+ }
381
+ else:
382
+ return output
383
+
384
+
385
+ class DatasetConditionedBiencoder(transformers.PreTrainedModel, ContextualModelMixin):
386
+ def __init__(
387
+ self,
388
+ config,
389
+ dataset_backbone: transformers.PreTrainedModel,
390
+ ):
391
+ super().__init__(config=config)
392
+ self.backbone = dataset_backbone
393
+ self.hidden_size = self.backbone.config.hidden_size
394
+ self.hidden_size = dataset_backbone.config.hidden_size
395
+ self.contextual_init()
396
+ self._shift_rotary_embedding()
397
+
398
+ self.pool_ignore_contextual_tokens = vars(self.config).get("pool_ignore_contextual_tokens", True)
399
+ self.pool_ignore_instruction_tokens = vars(self.config).get("pool_ignore_instruction_tokens", False)
400
+
401
+ tokenizer = transformers.AutoTokenizer.from_pretrained(self.config.embedder)
402
+ self.pool_instruction_end_id = tokenizer.encode(": ", add_special_tokens=False)[0] # Hardcoded for colon-ending prefixes.
403
+
404
+ @property
405
+ def num_corpus_tokens(self) -> int:
406
+ return self.config.transductive_corpus_size * self.transductive_tokens_per_document
407
+
408
+ def _shift_rotary_embedding(self) -> None:
409
+ disable_transductive_rotary_embedding = vars(self.config).get("disable_transductive_rotary_embedding", True)
410
+ if self.backbone.config.model_type.startswith("nomic") and disable_transductive_rotary_embedding:
411
+ # We only want to apply positional embeddings to the
412
+ # *text* portion of the backbone network.
413
+ self.backbone.config.rotary_start_pos = 0.0
414
+ rotary_disabled = 0
415
+
416
+ rotary_start_pos = self.num_corpus_tokens
417
+ for module in self.backbone.modules():
418
+ if hasattr(module, "rotary_emb_dim"):
419
+ module.rotary_start_pos = rotary_start_pos
420
+ rotary_disabled += 1
421
+ print0(f"modified {rotary_disabled} rotary modules – set rotary_start_pos to {rotary_start_pos}")
422
+
423
+ def forward(
424
+ self,
425
+ input_ids: torch.Tensor,
426
+ attention_mask: torch.Tensor,
427
+ dataset_embeddings: torch.Tensor,
428
+ output_hidden_states: bool = False,
429
+ null_dataset_embedding: bool = False,
430
+ ) -> torch.Tensor:
431
+ soft_prompt = self._prepare_dataset_embeddings(
432
+ input_ids=input_ids,
433
+ dataset_embeddings=dataset_embeddings,
434
+ null_dataset_embedding=null_dataset_embedding,
435
+ )
436
+ backbone_attention_mask = torch.ones(
437
+ soft_prompt.shape[0:2],
438
+ dtype=torch.long,
439
+ device=soft_prompt.device,
440
+ )
441
+ inputs_embeds = self.backbone.embeddings(input_ids) # (b, s) -> (b, s, d)
442
+ inputs_embeds = torch.cat((soft_prompt, inputs_embeds), dim=1) # (v, 4+b+s, d)
443
+ input_attention_mask = torch.cat((backbone_attention_mask, attention_mask), dim=1)
444
+ output = self.backbone(
445
+ inputs_embeds=inputs_embeds,
446
+ attention_mask=input_attention_mask,
447
+ ) # (1, 4 + b + s, d)
448
+ # trim soft prompt
449
+ output_vectors = output.last_hidden_state
450
+
451
+ # use only these tokens
452
+ n_soft_prompt_tokens = soft_prompt.shape[1]
453
+
454
+ if self.pool_ignore_instruction_tokens:
455
+ # Denote the end of an instruction with an extra BOS token.
456
+ # This is a bit arcane but relies on the fact that there will be a BOS token after the
457
+ # instruction, but also there may or may not be a BOS token at the beginning.
458
+ instruction_end_idx = (
459
+ (input_ids == self.pool_instruction_end_id) &
460
+ attention_mask &
461
+ (torch.arange(input_ids.shape[1], device=input_ids.device)[None, :] > 0)
462
+ ).int().argmax(1)
463
+ is_instruction_token_mask = (
464
+ torch.arange(input_ids.shape[1], device=input_ids.device)[None, :] <= instruction_end_idx[:, None]
465
+ )
466
+ # catch edge case where there is no instruction
467
+ is_instruction_token_mask = is_instruction_token_mask.where(
468
+ (instruction_end_idx > 0)[:, None], torch.zeros_like(is_instruction_token_mask)
469
+ )
470
+ output_attention_mask = torch.cat((backbone_attention_mask, attention_mask & ~is_instruction_token_mask), dim=1)
471
+ else:
472
+ output_attention_mask = input_attention_mask
473
+
474
+ if self.pool_ignore_contextual_tokens:
475
+ output_vectors = output_vectors[:, n_soft_prompt_tokens:, :]
476
+ output_attention_mask = output_attention_mask[:, n_soft_prompt_tokens:]
477
+ output_pooled = mean_pool(output_vectors, output_attention_mask)
478
+ # average with original vectors
479
+ output = self.output_projection(output_pooled) + output_pooled # (b, d) -> (b, d) / with residual connection
480
+
481
+ if output_hidden_states:
482
+ return {
483
+ "hidden_states": output_vectors,
484
+ "pooled": output,
485
+ }
486
+ else:
487
+ return output
488
+
489
+
490
+ class DatasetPrefixBiencoder(transformers.PreTrainedModel, ContextualModelMixin):
491
+ def __init__(
492
+ self,
493
+ config, #: transformers.PreTrainedConfig,
494
+ embedder: transformers.PreTrainedModel,
495
+ ):
496
+ super().__init__(config=config)
497
+ self.embedder = embedder
498
+ self.hidden_size = self.embedder.config.hidden_size
499
+ self.contextual_init()
500
+
501
+ def forward(
502
+ self,
503
+ input_ids: torch.Tensor,
504
+ attention_mask: torch.Tensor,
505
+ dataset_input_ids: torch.Tensor,
506
+ dataset_attention_mask: torch.Tensor,
507
+ output_hidden_states: bool = False,
508
+ ) -> torch.Tensor:
509
+ R = torch.randint(low=0, high=len(dataset_input_ids), size=(len(input_ids),), device=dataset_input_ids.device)
510
+
511
+ dataset_input_ids = dataset_input_ids[R]
512
+ input_ids = torch.cat((dataset_input_ids, input_ids), dim=1)
513
+
514
+ dataset_attention_mask = torch.ones_like(dataset_attention_mask, device=dataset_attention_mask.device)
515
+ input_attention_mask = torch.cat((dataset_attention_mask, attention_mask), dim=1)
516
+ output_attention_mask = torch.cat(
517
+ (torch.zeros_like(dataset_input_ids), attention_mask), dim=1
518
+ )
519
+
520
+ output = self.embedder(
521
+ input_ids=input_ids,
522
+ attention_mask=input_attention_mask,
523
+ )
524
+
525
+ output_vectors = output.last_hidden_state
526
+ output_pooled = mean_pool(output_vectors, output_attention_mask)
527
+ output = self.output_projection(output_pooled) # (b, 2d) -> (b, d)
528
+
529
+ if output_hidden_states:
530
+ S_d = dataset_attention_mask.shape[1]
531
+ output_vectors = output_vectors[:, S_d:, :]
532
+ return {
533
+ "hidden_states": output_vectors,
534
+ "pooled": output,
535
+ }
536
+ else:
537
+ return output
538
+
539
+
540
+ class ContextualDocumentEmbeddingTransformer(transformers.PreTrainedModel):
541
+ config_class = ContextualModelConfig
542
+ embedder: transformers.PreTrainedModel
543
+ dataset_backbone: transformers.PreTrainedModel
544
+ def __init__(
545
+ self,
546
+ config,
547
+ ):
548
+ super().__init__(config=config)
549
+ dataset_backbone, _ = load_embedder_and_tokenizer(
550
+ vars(config).get("dataset_backbone") or config.embedder
551
+ )
552
+
553
+ if config.limit_layers:
554
+ print0(f"Limiting layers to {config.limit_layers}")
555
+ limit_layers(dataset_backbone, config.limit_layers)
556
+
557
+ biencoder_config = copy.deepcopy(config)
558
+ biencoder_config.embedding_output_dim = None
559
+ biencoder_config.limit_layers = vars(self.config).get("limit_layers_first_stage", None)
560
+ self.first_stage_model = BiEncoder(
561
+ config=biencoder_config,
562
+ )
563
+
564
+ if vars(config).get("autoregressive_backbone", False):
565
+ self.second_stage_model = DatasetConditionedAutoregressive(
566
+ config=config,
567
+ dataset_backbone=dataset_backbone,
568
+ first_stage_hidden_size=self.first_stage_model.hidden_size,
569
+ )
570
+ else:
571
+ self.second_stage_model = DatasetConditionedBiencoder(
572
+ config=config,
573
+ dataset_backbone=dataset_backbone
574
+ )
575
+
576
+ self.temp = config.logit_scale
577
+ if config.disable_dropout:
578
+ disable_dropout(self)
579
+
580
+ transductive_tie_token_embeddings = vars(self.config).get("transductive_tie_token_embeddings", False)
581
+ if transductive_tie_token_embeddings:
582
+ self.second_stage_model.backbone.embeddings.word_embeddings.weight = (
583
+ self.first_stage_model.embedder.embeddings.word_embeddings.weight
584
+ )
585
+
586
+ def forward(
587
+ self,
588
+ input_ids: torch.Tensor,
589
+ attention_mask: torch.Tensor,
590
+ dataset_input_ids: Optional[torch.Tensor],
591
+ dataset_attention_mask: Optional[torch.Tensor],
592
+ output_hidden_states: bool = False,
593
+ ) -> torch.Tensor:
594
+ """
595
+ input_ids (long torch.Tensor) – ids of input tokens
596
+ attention_mask (bool torch.Tensor)
597
+ """
598
+ dataset_embeddings = self.first_stage_model(
599
+ input_ids=dataset_input_ids,
600
+ attention_mask=dataset_attention_mask
601
+ )
602
+ return self.second_stage_model(
603
+ input_ids=input_ids,
604
+ attention_mask=attention_mask,
605
+ dataset_embeddings=dataset_embeddings,
606
+ output_hidden_states=output_hidden_states,
607
+ )
608
+
609
+
610
+
611
+ def get_model_class(name: str):
612
+ if name in 'transductive':
613
+ return ContextualDocumentEmbeddingTransformer
614
+ elif name == 'biencoder':
615
+ return BiEncoder
616
+ elif name == "biencoder_plus_plus":
617
+ from cde.model_extra import BiEncoderPlusPlus
618
+ return BiEncoderPlusPlus
619
+ elif name == "dataset_prefix_biencoder":
620
+ return DatasetPrefixBiencoder
621
+ else:
622
+ raise ValueError(f'unknown model cls {name}')
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e7cca261c510de07c012f3019366f1b6c5720761b6966b0388faea6e70398983
3
+ size 1124594680