Commit
·
e2acffb
1
Parent(s):
7bd35e6
initial release
Browse files- README.md +31 -0
- config.json +501 -0
- configuration_modernbert.py +213 -0
- maker.py +97 -0
- modeling_modernbert.py +1351 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +51 -0
- tokenizer.json +0 -0
- tokenizer_config.json +62 -0
- training_args.bin +3 -0
- ud.py +112 -0
README.md
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- "ja"
|
| 4 |
+
tags:
|
| 5 |
+
- "japanese"
|
| 6 |
+
- "pos"
|
| 7 |
+
- "dependency-parsing"
|
| 8 |
+
- "modernbert"
|
| 9 |
+
base_model: KoichiYasuoka/modernbert-small-japanese-wikipedia-upos
|
| 10 |
+
datasets:
|
| 11 |
+
- "universal_dependencies"
|
| 12 |
+
license: "apache-2.0"
|
| 13 |
+
pipeline_tag: "token-classification"
|
| 14 |
+
widget:
|
| 15 |
+
- text: "全学年にわたって小学校の国語の教科書に挿し絵が用いられている"
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# modernbert-small-japanese-wikipedia-ud-square
|
| 19 |
+
|
| 20 |
+
## Model Description
|
| 21 |
+
|
| 22 |
+
This is a ModernBERT model pretrained for POS-tagging and dependency-parsing (using `goeswith` for subwords), derived from [modernbert-small-japanese-wikipedia-upos](https://huggingface.co/KoichiYasuoka/modernbert-small-japanese-wikipedia-upos) and [UD_Japanese-GSDLUW](https://github.com/UniversalDependencies/UD_Japanese-GSDLUW).
|
| 23 |
+
|
| 24 |
+
## How to Use
|
| 25 |
+
|
| 26 |
+
```py
|
| 27 |
+
from transformers import pipeline
|
| 28 |
+
nlp=pipeline("universal-dependencies","KoichiYasuoka/modernbert-small-japanese-wikipedia-ud-square",trust_remote_code=True,aggregation_strategy="simple")
|
| 29 |
+
print(nlp("全学年にわたって小学校の国語の教科書に挿し絵が用いられている"))
|
| 30 |
+
```
|
| 31 |
+
|
config.json
ADDED
|
@@ -0,0 +1,501 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"ModernBertForTokenClassification"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "configuration_modernbert.ModernBertConfig",
|
| 9 |
+
"AutoModel": "modeling_modernbert.ModernBertModel",
|
| 10 |
+
"AutoModelForMaskedLM": "modeling_modernbert.ModernBertForMaskedLM",
|
| 11 |
+
"AutoModelForSequenceClassification": "modeling_modernbert.ModernBertForSequenceClassification",
|
| 12 |
+
"AutoModelForTokenClassification": "modeling_modernbert.ModernBertForTokenClassification"
|
| 13 |
+
},
|
| 14 |
+
"bos_token_id": 0,
|
| 15 |
+
"classifier_activation": "gelu",
|
| 16 |
+
"classifier_bias": false,
|
| 17 |
+
"classifier_dropout": 0.0,
|
| 18 |
+
"classifier_pooling": "mean",
|
| 19 |
+
"cls_token_id": 0,
|
| 20 |
+
"custom_pipelines": {
|
| 21 |
+
"universal-dependencies": {
|
| 22 |
+
"impl": "ud.UniversalDependenciesPipeline",
|
| 23 |
+
"pt": "AutoModelForTokenClassification"
|
| 24 |
+
}
|
| 25 |
+
},
|
| 26 |
+
"decoder_bias": true,
|
| 27 |
+
"deterministic_flash_attn": false,
|
| 28 |
+
"embedding_dropout": 0.0,
|
| 29 |
+
"eos_token_id": 2,
|
| 30 |
+
"global_attn_every_n_layers": 3,
|
| 31 |
+
"global_rope_theta": 160000.0,
|
| 32 |
+
"gradient_checkpointing": false,
|
| 33 |
+
"hidden_activation": "gelu",
|
| 34 |
+
"hidden_size": 256,
|
| 35 |
+
"id2label": {
|
| 36 |
+
"0": "ADJ",
|
| 37 |
+
"1": "ADJ.",
|
| 38 |
+
"2": "ADJ.|[acl]",
|
| 39 |
+
"3": "ADJ.|[advcl]",
|
| 40 |
+
"4": "ADJ.|[amod]",
|
| 41 |
+
"5": "ADJ.|[ccomp]",
|
| 42 |
+
"6": "ADJ.|[csubj:outer]",
|
| 43 |
+
"7": "ADJ.|[csubj]",
|
| 44 |
+
"8": "ADJ.|[dep]",
|
| 45 |
+
"9": "ADJ.|[nmod]",
|
| 46 |
+
"10": "ADJ.|[nsubj]",
|
| 47 |
+
"11": "ADJ.|[obj]",
|
| 48 |
+
"12": "ADJ.|[obl]",
|
| 49 |
+
"13": "ADJ.|[root]",
|
| 50 |
+
"14": "ADJ|[acl]",
|
| 51 |
+
"15": "ADJ|[advcl]",
|
| 52 |
+
"16": "ADJ|[amod]",
|
| 53 |
+
"17": "ADJ|[ccomp]",
|
| 54 |
+
"18": "ADJ|[csubj:outer]",
|
| 55 |
+
"19": "ADJ|[csubj]",
|
| 56 |
+
"20": "ADJ|[dep]",
|
| 57 |
+
"21": "ADJ|[nmod]",
|
| 58 |
+
"22": "ADJ|[nsubj]",
|
| 59 |
+
"23": "ADJ|[obj]",
|
| 60 |
+
"24": "ADJ|[obl]",
|
| 61 |
+
"25": "ADJ|[root]",
|
| 62 |
+
"26": "ADP",
|
| 63 |
+
"27": "ADP.",
|
| 64 |
+
"28": "ADP.|[case]",
|
| 65 |
+
"29": "ADP.|[fixed]",
|
| 66 |
+
"30": "ADP|[case]",
|
| 67 |
+
"31": "ADP|[fixed]",
|
| 68 |
+
"32": "ADV",
|
| 69 |
+
"33": "ADV.",
|
| 70 |
+
"34": "ADV.|[advcl]",
|
| 71 |
+
"35": "ADV.|[advmod]",
|
| 72 |
+
"36": "ADV.|[dep]",
|
| 73 |
+
"37": "ADV.|[obj]",
|
| 74 |
+
"38": "ADV.|[root]",
|
| 75 |
+
"39": "ADV|[advcl]",
|
| 76 |
+
"40": "ADV|[advmod]",
|
| 77 |
+
"41": "ADV|[dep]",
|
| 78 |
+
"42": "ADV|[obj]",
|
| 79 |
+
"43": "ADV|[root]",
|
| 80 |
+
"44": "AUX",
|
| 81 |
+
"45": "AUX.",
|
| 82 |
+
"46": "AUX.|Polarity=Neg",
|
| 83 |
+
"47": "AUX.|Polarity=Neg|[aux]",
|
| 84 |
+
"48": "AUX.|Polarity=Neg|[fixed]",
|
| 85 |
+
"49": "AUX.|[aux]",
|
| 86 |
+
"50": "AUX.|[cop]",
|
| 87 |
+
"51": "AUX.|[fixed]",
|
| 88 |
+
"52": "AUX.|[root]",
|
| 89 |
+
"53": "AUX|Polarity=Neg",
|
| 90 |
+
"54": "AUX|Polarity=Neg|[aux]",
|
| 91 |
+
"55": "AUX|Polarity=Neg|[fixed]",
|
| 92 |
+
"56": "AUX|[aux]",
|
| 93 |
+
"57": "AUX|[cop]",
|
| 94 |
+
"58": "AUX|[fixed]",
|
| 95 |
+
"59": "AUX|[root]",
|
| 96 |
+
"60": "CCONJ",
|
| 97 |
+
"61": "CCONJ.",
|
| 98 |
+
"62": "CCONJ.|[cc]",
|
| 99 |
+
"63": "CCONJ|[cc]",
|
| 100 |
+
"64": "DET",
|
| 101 |
+
"65": "DET.",
|
| 102 |
+
"66": "DET.|[det]",
|
| 103 |
+
"67": "DET|[det]",
|
| 104 |
+
"68": "INTJ",
|
| 105 |
+
"69": "INTJ.",
|
| 106 |
+
"70": "INTJ.|[discourse]",
|
| 107 |
+
"71": "INTJ.|[root]",
|
| 108 |
+
"72": "INTJ|[discourse]",
|
| 109 |
+
"73": "INTJ|[root]",
|
| 110 |
+
"74": "NOUN",
|
| 111 |
+
"75": "NOUN.",
|
| 112 |
+
"76": "NOUN.|Polarity=Neg",
|
| 113 |
+
"77": "NOUN.|Polarity=Neg|[obl]",
|
| 114 |
+
"78": "NOUN.|Polarity=Neg|[root]",
|
| 115 |
+
"79": "NOUN.|[acl]",
|
| 116 |
+
"80": "NOUN.|[advcl]",
|
| 117 |
+
"81": "NOUN.|[ccomp]",
|
| 118 |
+
"82": "NOUN.|[compound]",
|
| 119 |
+
"83": "NOUN.|[csubj:outer]",
|
| 120 |
+
"84": "NOUN.|[csubj]",
|
| 121 |
+
"85": "NOUN.|[nmod]",
|
| 122 |
+
"86": "NOUN.|[nsubj:outer]",
|
| 123 |
+
"87": "NOUN.|[nsubj]",
|
| 124 |
+
"88": "NOUN.|[obj]",
|
| 125 |
+
"89": "NOUN.|[obl]",
|
| 126 |
+
"90": "NOUN.|[root]",
|
| 127 |
+
"91": "NOUN|Polarity=Neg",
|
| 128 |
+
"92": "NOUN|Polarity=Neg|[obl]",
|
| 129 |
+
"93": "NOUN|Polarity=Neg|[root]",
|
| 130 |
+
"94": "NOUN|[acl]",
|
| 131 |
+
"95": "NOUN|[advcl]",
|
| 132 |
+
"96": "NOUN|[ccomp]",
|
| 133 |
+
"97": "NOUN|[compound]",
|
| 134 |
+
"98": "NOUN|[csubj:outer]",
|
| 135 |
+
"99": "NOUN|[csubj]",
|
| 136 |
+
"100": "NOUN|[nmod]",
|
| 137 |
+
"101": "NOUN|[nsubj:outer]",
|
| 138 |
+
"102": "NOUN|[nsubj]",
|
| 139 |
+
"103": "NOUN|[obj]",
|
| 140 |
+
"104": "NOUN|[obl]",
|
| 141 |
+
"105": "NOUN|[root]",
|
| 142 |
+
"106": "NUM",
|
| 143 |
+
"107": "NUM.",
|
| 144 |
+
"108": "NUM.|[advcl]",
|
| 145 |
+
"109": "NUM.|[compound]",
|
| 146 |
+
"110": "NUM.|[nmod]",
|
| 147 |
+
"111": "NUM.|[nsubj:outer]",
|
| 148 |
+
"112": "NUM.|[nsubj]",
|
| 149 |
+
"113": "NUM.|[nummod]",
|
| 150 |
+
"114": "NUM.|[obj]",
|
| 151 |
+
"115": "NUM.|[obl]",
|
| 152 |
+
"116": "NUM.|[root]",
|
| 153 |
+
"117": "NUM|[advcl]",
|
| 154 |
+
"118": "NUM|[compound]",
|
| 155 |
+
"119": "NUM|[nmod]",
|
| 156 |
+
"120": "NUM|[nsubj:outer]",
|
| 157 |
+
"121": "NUM|[nsubj]",
|
| 158 |
+
"122": "NUM|[nummod]",
|
| 159 |
+
"123": "NUM|[obj]",
|
| 160 |
+
"124": "NUM|[obl]",
|
| 161 |
+
"125": "NUM|[root]",
|
| 162 |
+
"126": "PART",
|
| 163 |
+
"127": "PART.",
|
| 164 |
+
"128": "PART.|[mark]",
|
| 165 |
+
"129": "PART|[mark]",
|
| 166 |
+
"130": "PRON",
|
| 167 |
+
"131": "PRON.",
|
| 168 |
+
"132": "PRON.|[acl]",
|
| 169 |
+
"133": "PRON.|[advcl]",
|
| 170 |
+
"134": "PRON.|[nmod]",
|
| 171 |
+
"135": "PRON.|[nsubj:outer]",
|
| 172 |
+
"136": "PRON.|[nsubj]",
|
| 173 |
+
"137": "PRON.|[obj]",
|
| 174 |
+
"138": "PRON.|[obl]",
|
| 175 |
+
"139": "PRON.|[root]",
|
| 176 |
+
"140": "PRON|[acl]",
|
| 177 |
+
"141": "PRON|[advcl]",
|
| 178 |
+
"142": "PRON|[nmod]",
|
| 179 |
+
"143": "PRON|[nsubj:outer]",
|
| 180 |
+
"144": "PRON|[nsubj]",
|
| 181 |
+
"145": "PRON|[obj]",
|
| 182 |
+
"146": "PRON|[obl]",
|
| 183 |
+
"147": "PRON|[root]",
|
| 184 |
+
"148": "PROPN",
|
| 185 |
+
"149": "PROPN.",
|
| 186 |
+
"150": "PROPN.|[acl]",
|
| 187 |
+
"151": "PROPN.|[advcl]",
|
| 188 |
+
"152": "PROPN.|[compound]",
|
| 189 |
+
"153": "PROPN.|[nmod]",
|
| 190 |
+
"154": "PROPN.|[nsubj:outer]",
|
| 191 |
+
"155": "PROPN.|[nsubj]",
|
| 192 |
+
"156": "PROPN.|[obj]",
|
| 193 |
+
"157": "PROPN.|[obl]",
|
| 194 |
+
"158": "PROPN.|[root]",
|
| 195 |
+
"159": "PROPN|[acl]",
|
| 196 |
+
"160": "PROPN|[advcl]",
|
| 197 |
+
"161": "PROPN|[compound]",
|
| 198 |
+
"162": "PROPN|[nmod]",
|
| 199 |
+
"163": "PROPN|[nsubj:outer]",
|
| 200 |
+
"164": "PROPN|[nsubj]",
|
| 201 |
+
"165": "PROPN|[obj]",
|
| 202 |
+
"166": "PROPN|[obl]",
|
| 203 |
+
"167": "PROPN|[root]",
|
| 204 |
+
"168": "PUNCT",
|
| 205 |
+
"169": "PUNCT.",
|
| 206 |
+
"170": "PUNCT.|[punct]",
|
| 207 |
+
"171": "PUNCT|[punct]",
|
| 208 |
+
"172": "SCONJ",
|
| 209 |
+
"173": "SCONJ.",
|
| 210 |
+
"174": "SCONJ.|[dep]",
|
| 211 |
+
"175": "SCONJ.|[fixed]",
|
| 212 |
+
"176": "SCONJ.|[mark]",
|
| 213 |
+
"177": "SCONJ|[dep]",
|
| 214 |
+
"178": "SCONJ|[fixed]",
|
| 215 |
+
"179": "SCONJ|[mark]",
|
| 216 |
+
"180": "SYM",
|
| 217 |
+
"181": "SYM.",
|
| 218 |
+
"182": "SYM.|[compound]",
|
| 219 |
+
"183": "SYM.|[dep]",
|
| 220 |
+
"184": "SYM.|[nmod]",
|
| 221 |
+
"185": "SYM.|[obl]",
|
| 222 |
+
"186": "SYM|[compound]",
|
| 223 |
+
"187": "SYM|[dep]",
|
| 224 |
+
"188": "SYM|[nmod]",
|
| 225 |
+
"189": "SYM|[obl]",
|
| 226 |
+
"190": "VERB",
|
| 227 |
+
"191": "VERB.",
|
| 228 |
+
"192": "VERB.|[acl]",
|
| 229 |
+
"193": "VERB.|[advcl]",
|
| 230 |
+
"194": "VERB.|[ccomp]",
|
| 231 |
+
"195": "VERB.|[compound]",
|
| 232 |
+
"196": "VERB.|[csubj:outer]",
|
| 233 |
+
"197": "VERB.|[csubj]",
|
| 234 |
+
"198": "VERB.|[nmod]",
|
| 235 |
+
"199": "VERB.|[obj]",
|
| 236 |
+
"200": "VERB.|[obl]",
|
| 237 |
+
"201": "VERB.|[root]",
|
| 238 |
+
"202": "VERB|[acl]",
|
| 239 |
+
"203": "VERB|[advcl]",
|
| 240 |
+
"204": "VERB|[ccomp]",
|
| 241 |
+
"205": "VERB|[compound]",
|
| 242 |
+
"206": "VERB|[csubj:outer]",
|
| 243 |
+
"207": "VERB|[csubj]",
|
| 244 |
+
"208": "VERB|[nmod]",
|
| 245 |
+
"209": "VERB|[obj]",
|
| 246 |
+
"210": "VERB|[obl]",
|
| 247 |
+
"211": "VERB|[root]",
|
| 248 |
+
"212": "X",
|
| 249 |
+
"213": "X.",
|
| 250 |
+
"214": "X.|[dep]",
|
| 251 |
+
"215": "X.|[goeswith]",
|
| 252 |
+
"216": "X.|[nmod]",
|
| 253 |
+
"217": "X|[dep]",
|
| 254 |
+
"218": "X|[nmod]"
|
| 255 |
+
},
|
| 256 |
+
"initializer_cutoff_factor": 2.0,
|
| 257 |
+
"initializer_range": 0.02,
|
| 258 |
+
"intermediate_size": 768,
|
| 259 |
+
"label2id": {
|
| 260 |
+
"ADJ": 0,
|
| 261 |
+
"ADJ.": 1,
|
| 262 |
+
"ADJ.|[acl]": 2,
|
| 263 |
+
"ADJ.|[advcl]": 3,
|
| 264 |
+
"ADJ.|[amod]": 4,
|
| 265 |
+
"ADJ.|[ccomp]": 5,
|
| 266 |
+
"ADJ.|[csubj:outer]": 6,
|
| 267 |
+
"ADJ.|[csubj]": 7,
|
| 268 |
+
"ADJ.|[dep]": 8,
|
| 269 |
+
"ADJ.|[nmod]": 9,
|
| 270 |
+
"ADJ.|[nsubj]": 10,
|
| 271 |
+
"ADJ.|[obj]": 11,
|
| 272 |
+
"ADJ.|[obl]": 12,
|
| 273 |
+
"ADJ.|[root]": 13,
|
| 274 |
+
"ADJ|[acl]": 14,
|
| 275 |
+
"ADJ|[advcl]": 15,
|
| 276 |
+
"ADJ|[amod]": 16,
|
| 277 |
+
"ADJ|[ccomp]": 17,
|
| 278 |
+
"ADJ|[csubj:outer]": 18,
|
| 279 |
+
"ADJ|[csubj]": 19,
|
| 280 |
+
"ADJ|[dep]": 20,
|
| 281 |
+
"ADJ|[nmod]": 21,
|
| 282 |
+
"ADJ|[nsubj]": 22,
|
| 283 |
+
"ADJ|[obj]": 23,
|
| 284 |
+
"ADJ|[obl]": 24,
|
| 285 |
+
"ADJ|[root]": 25,
|
| 286 |
+
"ADP": 26,
|
| 287 |
+
"ADP.": 27,
|
| 288 |
+
"ADP.|[case]": 28,
|
| 289 |
+
"ADP.|[fixed]": 29,
|
| 290 |
+
"ADP|[case]": 30,
|
| 291 |
+
"ADP|[fixed]": 31,
|
| 292 |
+
"ADV": 32,
|
| 293 |
+
"ADV.": 33,
|
| 294 |
+
"ADV.|[advcl]": 34,
|
| 295 |
+
"ADV.|[advmod]": 35,
|
| 296 |
+
"ADV.|[dep]": 36,
|
| 297 |
+
"ADV.|[obj]": 37,
|
| 298 |
+
"ADV.|[root]": 38,
|
| 299 |
+
"ADV|[advcl]": 39,
|
| 300 |
+
"ADV|[advmod]": 40,
|
| 301 |
+
"ADV|[dep]": 41,
|
| 302 |
+
"ADV|[obj]": 42,
|
| 303 |
+
"ADV|[root]": 43,
|
| 304 |
+
"AUX": 44,
|
| 305 |
+
"AUX.": 45,
|
| 306 |
+
"AUX.|Polarity=Neg": 46,
|
| 307 |
+
"AUX.|Polarity=Neg|[aux]": 47,
|
| 308 |
+
"AUX.|Polarity=Neg|[fixed]": 48,
|
| 309 |
+
"AUX.|[aux]": 49,
|
| 310 |
+
"AUX.|[cop]": 50,
|
| 311 |
+
"AUX.|[fixed]": 51,
|
| 312 |
+
"AUX.|[root]": 52,
|
| 313 |
+
"AUX|Polarity=Neg": 53,
|
| 314 |
+
"AUX|Polarity=Neg|[aux]": 54,
|
| 315 |
+
"AUX|Polarity=Neg|[fixed]": 55,
|
| 316 |
+
"AUX|[aux]": 56,
|
| 317 |
+
"AUX|[cop]": 57,
|
| 318 |
+
"AUX|[fixed]": 58,
|
| 319 |
+
"AUX|[root]": 59,
|
| 320 |
+
"CCONJ": 60,
|
| 321 |
+
"CCONJ.": 61,
|
| 322 |
+
"CCONJ.|[cc]": 62,
|
| 323 |
+
"CCONJ|[cc]": 63,
|
| 324 |
+
"DET": 64,
|
| 325 |
+
"DET.": 65,
|
| 326 |
+
"DET.|[det]": 66,
|
| 327 |
+
"DET|[det]": 67,
|
| 328 |
+
"INTJ": 68,
|
| 329 |
+
"INTJ.": 69,
|
| 330 |
+
"INTJ.|[discourse]": 70,
|
| 331 |
+
"INTJ.|[root]": 71,
|
| 332 |
+
"INTJ|[discourse]": 72,
|
| 333 |
+
"INTJ|[root]": 73,
|
| 334 |
+
"NOUN": 74,
|
| 335 |
+
"NOUN.": 75,
|
| 336 |
+
"NOUN.|Polarity=Neg": 76,
|
| 337 |
+
"NOUN.|Polarity=Neg|[obl]": 77,
|
| 338 |
+
"NOUN.|Polarity=Neg|[root]": 78,
|
| 339 |
+
"NOUN.|[acl]": 79,
|
| 340 |
+
"NOUN.|[advcl]": 80,
|
| 341 |
+
"NOUN.|[ccomp]": 81,
|
| 342 |
+
"NOUN.|[compound]": 82,
|
| 343 |
+
"NOUN.|[csubj:outer]": 83,
|
| 344 |
+
"NOUN.|[csubj]": 84,
|
| 345 |
+
"NOUN.|[nmod]": 85,
|
| 346 |
+
"NOUN.|[nsubj:outer]": 86,
|
| 347 |
+
"NOUN.|[nsubj]": 87,
|
| 348 |
+
"NOUN.|[obj]": 88,
|
| 349 |
+
"NOUN.|[obl]": 89,
|
| 350 |
+
"NOUN.|[root]": 90,
|
| 351 |
+
"NOUN|Polarity=Neg": 91,
|
| 352 |
+
"NOUN|Polarity=Neg|[obl]": 92,
|
| 353 |
+
"NOUN|Polarity=Neg|[root]": 93,
|
| 354 |
+
"NOUN|[acl]": 94,
|
| 355 |
+
"NOUN|[advcl]": 95,
|
| 356 |
+
"NOUN|[ccomp]": 96,
|
| 357 |
+
"NOUN|[compound]": 97,
|
| 358 |
+
"NOUN|[csubj:outer]": 98,
|
| 359 |
+
"NOUN|[csubj]": 99,
|
| 360 |
+
"NOUN|[nmod]": 100,
|
| 361 |
+
"NOUN|[nsubj:outer]": 101,
|
| 362 |
+
"NOUN|[nsubj]": 102,
|
| 363 |
+
"NOUN|[obj]": 103,
|
| 364 |
+
"NOUN|[obl]": 104,
|
| 365 |
+
"NOUN|[root]": 105,
|
| 366 |
+
"NUM": 106,
|
| 367 |
+
"NUM.": 107,
|
| 368 |
+
"NUM.|[advcl]": 108,
|
| 369 |
+
"NUM.|[compound]": 109,
|
| 370 |
+
"NUM.|[nmod]": 110,
|
| 371 |
+
"NUM.|[nsubj:outer]": 111,
|
| 372 |
+
"NUM.|[nsubj]": 112,
|
| 373 |
+
"NUM.|[nummod]": 113,
|
| 374 |
+
"NUM.|[obj]": 114,
|
| 375 |
+
"NUM.|[obl]": 115,
|
| 376 |
+
"NUM.|[root]": 116,
|
| 377 |
+
"NUM|[advcl]": 117,
|
| 378 |
+
"NUM|[compound]": 118,
|
| 379 |
+
"NUM|[nmod]": 119,
|
| 380 |
+
"NUM|[nsubj:outer]": 120,
|
| 381 |
+
"NUM|[nsubj]": 121,
|
| 382 |
+
"NUM|[nummod]": 122,
|
| 383 |
+
"NUM|[obj]": 123,
|
| 384 |
+
"NUM|[obl]": 124,
|
| 385 |
+
"NUM|[root]": 125,
|
| 386 |
+
"PART": 126,
|
| 387 |
+
"PART.": 127,
|
| 388 |
+
"PART.|[mark]": 128,
|
| 389 |
+
"PART|[mark]": 129,
|
| 390 |
+
"PRON": 130,
|
| 391 |
+
"PRON.": 131,
|
| 392 |
+
"PRON.|[acl]": 132,
|
| 393 |
+
"PRON.|[advcl]": 133,
|
| 394 |
+
"PRON.|[nmod]": 134,
|
| 395 |
+
"PRON.|[nsubj:outer]": 135,
|
| 396 |
+
"PRON.|[nsubj]": 136,
|
| 397 |
+
"PRON.|[obj]": 137,
|
| 398 |
+
"PRON.|[obl]": 138,
|
| 399 |
+
"PRON.|[root]": 139,
|
| 400 |
+
"PRON|[acl]": 140,
|
| 401 |
+
"PRON|[advcl]": 141,
|
| 402 |
+
"PRON|[nmod]": 142,
|
| 403 |
+
"PRON|[nsubj:outer]": 143,
|
| 404 |
+
"PRON|[nsubj]": 144,
|
| 405 |
+
"PRON|[obj]": 145,
|
| 406 |
+
"PRON|[obl]": 146,
|
| 407 |
+
"PRON|[root]": 147,
|
| 408 |
+
"PROPN": 148,
|
| 409 |
+
"PROPN.": 149,
|
| 410 |
+
"PROPN.|[acl]": 150,
|
| 411 |
+
"PROPN.|[advcl]": 151,
|
| 412 |
+
"PROPN.|[compound]": 152,
|
| 413 |
+
"PROPN.|[nmod]": 153,
|
| 414 |
+
"PROPN.|[nsubj:outer]": 154,
|
| 415 |
+
"PROPN.|[nsubj]": 155,
|
| 416 |
+
"PROPN.|[obj]": 156,
|
| 417 |
+
"PROPN.|[obl]": 157,
|
| 418 |
+
"PROPN.|[root]": 158,
|
| 419 |
+
"PROPN|[acl]": 159,
|
| 420 |
+
"PROPN|[advcl]": 160,
|
| 421 |
+
"PROPN|[compound]": 161,
|
| 422 |
+
"PROPN|[nmod]": 162,
|
| 423 |
+
"PROPN|[nsubj:outer]": 163,
|
| 424 |
+
"PROPN|[nsubj]": 164,
|
| 425 |
+
"PROPN|[obj]": 165,
|
| 426 |
+
"PROPN|[obl]": 166,
|
| 427 |
+
"PROPN|[root]": 167,
|
| 428 |
+
"PUNCT": 168,
|
| 429 |
+
"PUNCT.": 169,
|
| 430 |
+
"PUNCT.|[punct]": 170,
|
| 431 |
+
"PUNCT|[punct]": 171,
|
| 432 |
+
"SCONJ": 172,
|
| 433 |
+
"SCONJ.": 173,
|
| 434 |
+
"SCONJ.|[dep]": 174,
|
| 435 |
+
"SCONJ.|[fixed]": 175,
|
| 436 |
+
"SCONJ.|[mark]": 176,
|
| 437 |
+
"SCONJ|[dep]": 177,
|
| 438 |
+
"SCONJ|[fixed]": 178,
|
| 439 |
+
"SCONJ|[mark]": 179,
|
| 440 |
+
"SYM": 180,
|
| 441 |
+
"SYM.": 181,
|
| 442 |
+
"SYM.|[compound]": 182,
|
| 443 |
+
"SYM.|[dep]": 183,
|
| 444 |
+
"SYM.|[nmod]": 184,
|
| 445 |
+
"SYM.|[obl]": 185,
|
| 446 |
+
"SYM|[compound]": 186,
|
| 447 |
+
"SYM|[dep]": 187,
|
| 448 |
+
"SYM|[nmod]": 188,
|
| 449 |
+
"SYM|[obl]": 189,
|
| 450 |
+
"VERB": 190,
|
| 451 |
+
"VERB.": 191,
|
| 452 |
+
"VERB.|[acl]": 192,
|
| 453 |
+
"VERB.|[advcl]": 193,
|
| 454 |
+
"VERB.|[ccomp]": 194,
|
| 455 |
+
"VERB.|[compound]": 195,
|
| 456 |
+
"VERB.|[csubj:outer]": 196,
|
| 457 |
+
"VERB.|[csubj]": 197,
|
| 458 |
+
"VERB.|[nmod]": 198,
|
| 459 |
+
"VERB.|[obj]": 199,
|
| 460 |
+
"VERB.|[obl]": 200,
|
| 461 |
+
"VERB.|[root]": 201,
|
| 462 |
+
"VERB|[acl]": 202,
|
| 463 |
+
"VERB|[advcl]": 203,
|
| 464 |
+
"VERB|[ccomp]": 204,
|
| 465 |
+
"VERB|[compound]": 205,
|
| 466 |
+
"VERB|[csubj:outer]": 206,
|
| 467 |
+
"VERB|[csubj]": 207,
|
| 468 |
+
"VERB|[nmod]": 208,
|
| 469 |
+
"VERB|[obj]": 209,
|
| 470 |
+
"VERB|[obl]": 210,
|
| 471 |
+
"VERB|[root]": 211,
|
| 472 |
+
"X": 212,
|
| 473 |
+
"X.": 213,
|
| 474 |
+
"X.|[dep]": 214,
|
| 475 |
+
"X.|[goeswith]": 215,
|
| 476 |
+
"X.|[nmod]": 216,
|
| 477 |
+
"X|[dep]": 217,
|
| 478 |
+
"X|[nmod]": 218
|
| 479 |
+
},
|
| 480 |
+
"layer_norm_eps": 1e-05,
|
| 481 |
+
"local_attention": 128,
|
| 482 |
+
"local_rope_theta": 10000.0,
|
| 483 |
+
"max_position_embeddings": 8192,
|
| 484 |
+
"mlp_bias": false,
|
| 485 |
+
"mlp_dropout": 0.0,
|
| 486 |
+
"model_type": "modernbert",
|
| 487 |
+
"norm_bias": false,
|
| 488 |
+
"norm_eps": 1e-05,
|
| 489 |
+
"num_attention_heads": 8,
|
| 490 |
+
"num_hidden_layers": 16,
|
| 491 |
+
"pad_token_id": 1,
|
| 492 |
+
"position_embedding_type": "absolute",
|
| 493 |
+
"reference_compile": true,
|
| 494 |
+
"sep_token_id": 2,
|
| 495 |
+
"sparse_pred_ignore_index": -100,
|
| 496 |
+
"sparse_prediction": false,
|
| 497 |
+
"tokenizer_class": "DebertaV2TokenizerFast",
|
| 498 |
+
"torch_dtype": "float32",
|
| 499 |
+
"transformers_version": "4.47.1",
|
| 500 |
+
"vocab_size": 65000
|
| 501 |
+
}
|
configuration_modernbert.py
ADDED
|
@@ -0,0 +1,213 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 2 |
+
# This file was automatically generated from src/transformers/models/modernbert/modular_modernbert.py.
|
| 3 |
+
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
| 4 |
+
# the file from the modular. If any change should be done, please apply the change to the
|
| 5 |
+
# modular_modernbert.py file directly. One of our CI enforces this.
|
| 6 |
+
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 7 |
+
# Copyright 2024 Answer.AI, LightOn, and contributors, and the HuggingFace Inc. team. All rights reserved.
|
| 8 |
+
#
|
| 9 |
+
#
|
| 10 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 11 |
+
# you may not use this file except in compliance with the License.
|
| 12 |
+
# You may obtain a copy of the License at
|
| 13 |
+
#
|
| 14 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 15 |
+
#
|
| 16 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 17 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 18 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 19 |
+
# See the License for the specific language governing permissions and
|
| 20 |
+
# limitations under the License.
|
| 21 |
+
|
| 22 |
+
from typing import Literal
|
| 23 |
+
|
| 24 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class ModernBertConfig(PretrainedConfig):
|
| 28 |
+
r"""
|
| 29 |
+
This is the configuration class to store the configuration of a [`ModernBertModel`]. It is used to instantiate an ModernBert
|
| 30 |
+
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
| 31 |
+
defaults will yield a similar configuration to that of the ModernBERT-base.
|
| 32 |
+
e.g. [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base)
|
| 33 |
+
|
| 34 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 35 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 36 |
+
|
| 37 |
+
Args:
|
| 38 |
+
vocab_size (`int`, *optional*, defaults to 50368):
|
| 39 |
+
Vocabulary size of the ModernBert model. Defines the number of different tokens that can be represented by the
|
| 40 |
+
`inputs_ids` passed when calling [`ModernBertModel`]
|
| 41 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
| 42 |
+
Dimension of the hidden representations.
|
| 43 |
+
intermediate_size (`int`, *optional*, defaults to 1152):
|
| 44 |
+
Dimension of the MLP representations.
|
| 45 |
+
num_hidden_layers (`int`, *optional*, defaults to 22):
|
| 46 |
+
Number of hidden layers in the Transformer decoder.
|
| 47 |
+
num_attention_heads (`int`, *optional*, defaults to 12):
|
| 48 |
+
Number of attention heads for each attention layer in the Transformer decoder.
|
| 49 |
+
hidden_activation (`str` or `function`, *optional*, defaults to `"gelu"`):
|
| 50 |
+
The non-linear activation function (function or string) in the decoder. Will default to `"gelu"`
|
| 51 |
+
if not specified.
|
| 52 |
+
max_position_embeddings (`int`, *optional*, defaults to 8192):
|
| 53 |
+
The maximum sequence length that this model might ever be used with.
|
| 54 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 55 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 56 |
+
initializer_cutoff_factor (`float`, *optional*, defaults to 2.0):
|
| 57 |
+
The cutoff factor for the truncated_normal_initializer for initializing all weight matrices.
|
| 58 |
+
norm_eps (`float`, *optional*, defaults to 1e-05):
|
| 59 |
+
The epsilon used by the rms normalization layers.
|
| 60 |
+
norm_bias (`bool`, *optional*, defaults to `False`):
|
| 61 |
+
Whether to use bias in the normalization layers.
|
| 62 |
+
pad_token_id (`int`, *optional*, defaults to 50283):
|
| 63 |
+
Padding token id.
|
| 64 |
+
eos_token_id (`int`, *optional*, defaults to 50282):
|
| 65 |
+
End of stream token id.
|
| 66 |
+
bos_token_id (`int`, *optional*, defaults to 50281):
|
| 67 |
+
Beginning of stream token id.
|
| 68 |
+
cls_token_id (`int`, *optional*, defaults to 50281):
|
| 69 |
+
Classification token id.
|
| 70 |
+
sep_token_id (`int`, *optional*, defaults to 50282):
|
| 71 |
+
Separation token id.
|
| 72 |
+
global_rope_theta (`float`, *optional*, defaults to 160000.0):
|
| 73 |
+
The base period of the global RoPE embeddings.
|
| 74 |
+
attention_bias (`bool`, *optional*, defaults to `False`):
|
| 75 |
+
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
| 76 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 77 |
+
The dropout ratio for the attention probabilities.
|
| 78 |
+
global_attn_every_n_layers (`int`, *optional*, defaults to 3):
|
| 79 |
+
The number of layers between global attention layers.
|
| 80 |
+
local_attention (`int`, *optional*, defaults to 128):
|
| 81 |
+
The window size for local attention.
|
| 82 |
+
local_rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 83 |
+
The base period of the local RoPE embeddings.
|
| 84 |
+
embedding_dropout (`float`, *optional*, defaults to 0.0):
|
| 85 |
+
The dropout ratio for the embeddings.
|
| 86 |
+
mlp_bias (`bool`, *optional*, defaults to `False`):
|
| 87 |
+
Whether to use bias in the MLP layers.
|
| 88 |
+
mlp_dropout (`float`, *optional*, defaults to 0.0):
|
| 89 |
+
The dropout ratio for the MLP layers.
|
| 90 |
+
decoder_bias (`bool`, *optional*, defaults to `True`):
|
| 91 |
+
Whether to use bias in the decoder layers.
|
| 92 |
+
classifier_pooling (`str`, *optional*, defaults to `"cls"`):
|
| 93 |
+
The pooling method for the classifier. Should be either `"cls"` or `"mean"`. In local attention layers, the
|
| 94 |
+
CLS token doesn't attend to all tokens on long sequences.
|
| 95 |
+
classifier_dropout (`float`, *optional*, defaults to 0.0):
|
| 96 |
+
The dropout ratio for the classifier.
|
| 97 |
+
classifier_bias (`bool`, *optional*, defaults to `False`):
|
| 98 |
+
Whether to use bias in the classifier.
|
| 99 |
+
classifier_activation (`str`, *optional*, defaults to `"gelu"`):
|
| 100 |
+
The activation function for the classifier.
|
| 101 |
+
deterministic_flash_attn (`bool`, *optional*, defaults to `False`):
|
| 102 |
+
Whether to use deterministic flash attention. If `False`, inference will be faster but not deterministic.
|
| 103 |
+
sparse_prediction (`bool`, *optional*, defaults to `False`):
|
| 104 |
+
Whether to use sparse prediction for the masked language model instead of returning the full dense logits.
|
| 105 |
+
sparse_pred_ignore_index (`int`, *optional*, defaults to -100):
|
| 106 |
+
The index to ignore for the sparse prediction.
|
| 107 |
+
reference_compile (`bool`, *optional*):
|
| 108 |
+
Whether to compile the layers of the model which were compiled during pretraining. If `None`, then parts of
|
| 109 |
+
the model will be compiled if 1) `triton` is installed, 2) the model is not on MPS, 3) the model is not
|
| 110 |
+
shared between devices, and 4) the model is not resized after initialization. If `True`, then the model may
|
| 111 |
+
be faster in some scenarios.
|
| 112 |
+
|
| 113 |
+
Examples:
|
| 114 |
+
|
| 115 |
+
```python
|
| 116 |
+
>>> from transformers import ModernBertModel, ModernBertConfig
|
| 117 |
+
|
| 118 |
+
>>> # Initializing a ModernBert style configuration
|
| 119 |
+
>>> configuration = ModernBertConfig()
|
| 120 |
+
|
| 121 |
+
>>> # Initializing a model from the modernbert-base style configuration
|
| 122 |
+
>>> model = ModernBertModel(configuration)
|
| 123 |
+
|
| 124 |
+
>>> # Accessing the model configuration
|
| 125 |
+
>>> configuration = model.config
|
| 126 |
+
```"""
|
| 127 |
+
|
| 128 |
+
model_type = "modernbert"
|
| 129 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 130 |
+
|
| 131 |
+
def __init__(
|
| 132 |
+
self,
|
| 133 |
+
vocab_size=50368,
|
| 134 |
+
hidden_size=768,
|
| 135 |
+
intermediate_size=1152,
|
| 136 |
+
num_hidden_layers=22,
|
| 137 |
+
num_attention_heads=12,
|
| 138 |
+
hidden_activation="gelu",
|
| 139 |
+
max_position_embeddings=8192,
|
| 140 |
+
initializer_range=0.02,
|
| 141 |
+
initializer_cutoff_factor=2.0,
|
| 142 |
+
norm_eps=1e-5,
|
| 143 |
+
norm_bias=False,
|
| 144 |
+
pad_token_id=50283,
|
| 145 |
+
eos_token_id=50282,
|
| 146 |
+
bos_token_id=50281,
|
| 147 |
+
cls_token_id=50281,
|
| 148 |
+
sep_token_id=50282,
|
| 149 |
+
global_rope_theta=160000.0,
|
| 150 |
+
attention_bias=False,
|
| 151 |
+
attention_dropout=0.0,
|
| 152 |
+
global_attn_every_n_layers=3,
|
| 153 |
+
local_attention=128,
|
| 154 |
+
local_rope_theta=10000.0,
|
| 155 |
+
embedding_dropout=0.0,
|
| 156 |
+
mlp_bias=False,
|
| 157 |
+
mlp_dropout=0.0,
|
| 158 |
+
decoder_bias=True,
|
| 159 |
+
classifier_pooling: Literal["cls", "mean"] = "cls",
|
| 160 |
+
classifier_dropout=0.0,
|
| 161 |
+
classifier_bias=False,
|
| 162 |
+
classifier_activation="gelu",
|
| 163 |
+
deterministic_flash_attn=False,
|
| 164 |
+
sparse_prediction=False,
|
| 165 |
+
sparse_pred_ignore_index=-100,
|
| 166 |
+
reference_compile=None,
|
| 167 |
+
**kwargs,
|
| 168 |
+
):
|
| 169 |
+
super().__init__(
|
| 170 |
+
pad_token_id=pad_token_id,
|
| 171 |
+
bos_token_id=bos_token_id,
|
| 172 |
+
eos_token_id=eos_token_id,
|
| 173 |
+
cls_token_id=cls_token_id,
|
| 174 |
+
sep_token_id=sep_token_id,
|
| 175 |
+
**kwargs,
|
| 176 |
+
)
|
| 177 |
+
self.vocab_size = vocab_size
|
| 178 |
+
self.max_position_embeddings = max_position_embeddings
|
| 179 |
+
self.hidden_size = hidden_size
|
| 180 |
+
self.intermediate_size = intermediate_size
|
| 181 |
+
self.num_hidden_layers = num_hidden_layers
|
| 182 |
+
self.num_attention_heads = num_attention_heads
|
| 183 |
+
self.initializer_range = initializer_range
|
| 184 |
+
self.initializer_cutoff_factor = initializer_cutoff_factor
|
| 185 |
+
self.norm_eps = norm_eps
|
| 186 |
+
self.norm_bias = norm_bias
|
| 187 |
+
self.global_rope_theta = global_rope_theta
|
| 188 |
+
self.attention_bias = attention_bias
|
| 189 |
+
self.attention_dropout = attention_dropout
|
| 190 |
+
self.hidden_activation = hidden_activation
|
| 191 |
+
self.global_attn_every_n_layers = global_attn_every_n_layers
|
| 192 |
+
self.local_attention = local_attention
|
| 193 |
+
self.local_rope_theta = local_rope_theta
|
| 194 |
+
self.embedding_dropout = embedding_dropout
|
| 195 |
+
self.mlp_bias = mlp_bias
|
| 196 |
+
self.mlp_dropout = mlp_dropout
|
| 197 |
+
self.decoder_bias = decoder_bias
|
| 198 |
+
self.classifier_pooling = classifier_pooling
|
| 199 |
+
self.classifier_dropout = classifier_dropout
|
| 200 |
+
self.classifier_bias = classifier_bias
|
| 201 |
+
self.classifier_activation = classifier_activation
|
| 202 |
+
self.deterministic_flash_attn = deterministic_flash_attn
|
| 203 |
+
self.sparse_prediction = sparse_prediction
|
| 204 |
+
self.sparse_pred_ignore_index = sparse_pred_ignore_index
|
| 205 |
+
self.reference_compile = reference_compile
|
| 206 |
+
|
| 207 |
+
if self.classifier_pooling not in ["cls", "mean"]:
|
| 208 |
+
raise ValueError(
|
| 209 |
+
f'Invalid value for `classifier_pooling`, should be either "cls" or "mean", but is {self.classifier_pooling}.'
|
| 210 |
+
)
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
__all__ = ["ModernBertConfig"]
|
maker.py
ADDED
|
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#! /usr/bin/python3
|
| 2 |
+
src="KoichiYasuoka/modernbert-small-japanese-wikipedia-upos"
|
| 3 |
+
tgt="KoichiYasuoka/modernbert-small-japanese-wikipedia-ud-square"
|
| 4 |
+
url="https://github.com/UniversalDependencies/UD_Japanese-GSDLUW"
|
| 5 |
+
import os,numpy
|
| 6 |
+
d=os.path.basename(url)
|
| 7 |
+
os.system("test -d "+d+" || git clone --depth=1 "+url)
|
| 8 |
+
os.system("for F in train dev test ; do cp "+d+"/*-$F.conllu $F.conllu ; done")
|
| 9 |
+
class UDSquareDataset(object):
|
| 10 |
+
def __init__(self,conllu,tokenizer):
|
| 11 |
+
self.conllu=open(conllu,"r",encoding="utf-8")
|
| 12 |
+
self.tokenizer=tokenizer
|
| 13 |
+
self.seeks=[0]
|
| 14 |
+
label=set(["SYM.","X.","X.|[goeswith]"])
|
| 15 |
+
s=self.conllu.readline()
|
| 16 |
+
while s!="":
|
| 17 |
+
if s=="\n":
|
| 18 |
+
self.seeks.append(self.conllu.tell())
|
| 19 |
+
else:
|
| 20 |
+
w=s.split("\t")
|
| 21 |
+
if len(w)==10:
|
| 22 |
+
if w[0].isdecimal():
|
| 23 |
+
p=w[3]
|
| 24 |
+
q="" if w[5]=="_" else "|"+w[5]
|
| 25 |
+
r="|["+w[7]+"]"
|
| 26 |
+
label.add(p+q)
|
| 27 |
+
label.add(p+"."+q)
|
| 28 |
+
label.add(p+q+r)
|
| 29 |
+
label.add(p+"."+q+r)
|
| 30 |
+
s=self.conllu.readline()
|
| 31 |
+
self.label2id={l:i for i,l in enumerate(sorted(label))}
|
| 32 |
+
def __call__(*args):
|
| 33 |
+
lid={l:i for i,l in enumerate(sorted(set(sum([list(t.label2id) for t in args],[]))))}
|
| 34 |
+
for t in args:
|
| 35 |
+
t.label2id=lid
|
| 36 |
+
return lid
|
| 37 |
+
def __del__(self):
|
| 38 |
+
self.conllu.close()
|
| 39 |
+
__len__=lambda self:len(self.seeks)-1
|
| 40 |
+
def __getitem__(self,i):
|
| 41 |
+
self.conllu.seek(self.seeks[i])
|
| 42 |
+
c,t=[],[""]
|
| 43 |
+
while t[0]!="\n":
|
| 44 |
+
t=self.conllu.readline().split("\t")
|
| 45 |
+
if len(t)==10 and t[0].isdecimal():
|
| 46 |
+
c.append(t)
|
| 47 |
+
h={t[6] for t in c}
|
| 48 |
+
for t in c:
|
| 49 |
+
if t[6]!="0" and t[0] not in h:
|
| 50 |
+
t[3]+="."
|
| 51 |
+
v=self.tokenizer([t[1] for t in c],add_special_tokens=False)["input_ids"]
|
| 52 |
+
for i in range(len(v)-1,-1,-1):
|
| 53 |
+
if v[i]==[]:
|
| 54 |
+
v[i]=[self.tokenizer.unk_token_id]
|
| 55 |
+
if len(v[i])>1:
|
| 56 |
+
c[i][3]=c[i][3].replace(".","")
|
| 57 |
+
for j in range(1,len(v[i])):
|
| 58 |
+
c.insert(i+1,[c[i][0],"_","_","X.","_","_",c[i][0],"goeswith","_","_"])
|
| 59 |
+
y=["0"]+[t[0] for t in c]
|
| 60 |
+
h=[i if t[6]=="0" else y.index(t[6]) for i,t in enumerate(c,1)]
|
| 61 |
+
p=[t[3] if t[5]=="_" else t[3]+"|"+t[5] for t in c]
|
| 62 |
+
d=["|["+t[7]+"]" for t in c]
|
| 63 |
+
x=[not t[3].endswith(".") for t in c]
|
| 64 |
+
if len(x)<90:
|
| 65 |
+
x=[True]*len(x)
|
| 66 |
+
else:
|
| 67 |
+
w=(sum([1 for b in x if b])+1)*(len(x)+1)+1
|
| 68 |
+
for i in numpy.argsort([-abs(j-i-1) for i,j in enumerate(h)]):
|
| 69 |
+
if w+len(x)>8191:
|
| 70 |
+
break
|
| 71 |
+
if not x[i]:
|
| 72 |
+
x[i]=True
|
| 73 |
+
w+=len(x)+1
|
| 74 |
+
v=sum(v,[])
|
| 75 |
+
ids=[self.tokenizer.cls_token_id]+v+[self.tokenizer.sep_token_id]
|
| 76 |
+
upos=["SYM."]+p+["SYM."]
|
| 77 |
+
for i in range(len(v)):
|
| 78 |
+
if x[i]:
|
| 79 |
+
for j in range(len(v)):
|
| 80 |
+
ids.append(self.tokenizer.mask_token_id if i==j else v[j])
|
| 81 |
+
upos.append(p[j]+d[j] if h[j]==i+1 else p[j])
|
| 82 |
+
ids.append(self.tokenizer.sep_token_id)
|
| 83 |
+
upos.append("SYM.")
|
| 84 |
+
return {"input_ids":ids,"labels":[self.label2id[p] for p in upos]}
|
| 85 |
+
from transformers import AutoTokenizer,AutoConfig,AutoModelForTokenClassification,DataCollatorForTokenClassification,TrainingArguments,Trainer
|
| 86 |
+
tkz=AutoTokenizer.from_pretrained(src)
|
| 87 |
+
trainDS=UDSquareDataset("train.conllu",tkz)
|
| 88 |
+
devDS=UDSquareDataset("dev.conllu",tkz)
|
| 89 |
+
testDS=UDSquareDataset("test.conllu",tkz)
|
| 90 |
+
lid=trainDS(devDS,testDS)
|
| 91 |
+
cfg=AutoConfig.from_pretrained(src,num_labels=len(lid),label2id=lid,id2label={i:l for l,i in lid.items()},ignore_mismatched_sizes=True,trust_remote_code=True)
|
| 92 |
+
mdl=AutoModelForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True,trust_remote_code=True)
|
| 93 |
+
arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=1,dataloader_pin_memory=False,output_dir=tgt,overwrite_output_dir=True,save_total_limit=2,learning_rate=5e-05,warmup_ratio=0.1,save_safetensors=False)
|
| 94 |
+
trn=Trainer(args=arg,data_collator=DataCollatorForTokenClassification(tkz),model=mdl,train_dataset=trainDS)
|
| 95 |
+
trn.train()
|
| 96 |
+
trn.save_model(tgt)
|
| 97 |
+
tkz.save_pretrained(tgt)
|
modeling_modernbert.py
ADDED
|
@@ -0,0 +1,1351 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 2 |
+
# This file was automatically generated from src/transformers/models/modernbert/modular_modernbert.py.
|
| 3 |
+
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
| 4 |
+
# the file from the modular. If any change should be done, please apply the change to the
|
| 5 |
+
# modular_modernbert.py file directly. One of our CI enforces this.
|
| 6 |
+
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 7 |
+
# Copyright 2024 Answer.AI, LightOn, and contributors, and the HuggingFace Inc. team. All rights reserved.
|
| 8 |
+
#
|
| 9 |
+
#
|
| 10 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 11 |
+
# you may not use this file except in compliance with the License.
|
| 12 |
+
# You may obtain a copy of the License at
|
| 13 |
+
#
|
| 14 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 15 |
+
#
|
| 16 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 17 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 18 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 19 |
+
# See the License for the specific language governing permissions and
|
| 20 |
+
# limitations under the License.
|
| 21 |
+
|
| 22 |
+
import math
|
| 23 |
+
from typing import Dict, Optional, Tuple, Union
|
| 24 |
+
|
| 25 |
+
import torch
|
| 26 |
+
import torch.nn.functional as F
|
| 27 |
+
from torch import nn
|
| 28 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 29 |
+
|
| 30 |
+
from transformers.activations import ACT2FN
|
| 31 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
|
| 32 |
+
from transformers.modeling_outputs import BaseModelOutput, MaskedLMOutput, SequenceClassifierOutput, TokenClassifierOutput
|
| 33 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 34 |
+
from transformers.utils import (
|
| 35 |
+
add_code_sample_docstrings,
|
| 36 |
+
add_start_docstrings,
|
| 37 |
+
add_start_docstrings_to_model_forward,
|
| 38 |
+
is_flash_attn_2_available,
|
| 39 |
+
logging,
|
| 40 |
+
)
|
| 41 |
+
import importlib
|
| 42 |
+
is_triton_available = lambda: importlib.util.find_spec("triton") is not None
|
| 43 |
+
from .configuration_modernbert import ModernBertConfig
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
if is_flash_attn_2_available():
|
| 47 |
+
from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
|
| 48 |
+
from flash_attn.layers.rotary import RotaryEmbedding
|
| 49 |
+
from flash_attn.ops.triton.rotary import apply_rotary
|
| 50 |
+
else:
|
| 51 |
+
RotaryEmbedding = object
|
| 52 |
+
|
| 53 |
+
logger = logging.get_logger(__name__)
|
| 54 |
+
|
| 55 |
+
_CHECKPOINT_FOR_DOC = "answerdotai/ModernBERT-base"
|
| 56 |
+
_CONFIG_FOR_DOC = "ModernBertConfig"
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
class ApplyRotaryEmbUnpad(torch.autograd.Function):
|
| 60 |
+
@staticmethod
|
| 61 |
+
def forward(
|
| 62 |
+
ctx,
|
| 63 |
+
qkv,
|
| 64 |
+
cos,
|
| 65 |
+
sin,
|
| 66 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 67 |
+
max_seqlen: Optional[int] = None,
|
| 68 |
+
):
|
| 69 |
+
# (total_nnz, 3, nheads, headdim)
|
| 70 |
+
qkv = qkv.contiguous()
|
| 71 |
+
total_nnz, _three, _nheads, headdim = qkv.shape
|
| 72 |
+
# We need qkv to be contiguous so that when we reshape to combine (3, nheads) dimensions,
|
| 73 |
+
# we get the same tensor
|
| 74 |
+
# qk = rearrange(qkv[:, :2], "b_s t h d -> b_s (t h) d")
|
| 75 |
+
qk = qkv[:, :2].view(total_nnz, -1, headdim)
|
| 76 |
+
apply_rotary(
|
| 77 |
+
qk,
|
| 78 |
+
cos,
|
| 79 |
+
sin,
|
| 80 |
+
seqlen_offsets=0,
|
| 81 |
+
cu_seqlens=cu_seqlens,
|
| 82 |
+
max_seqlen=max_seqlen,
|
| 83 |
+
interleaved=False,
|
| 84 |
+
inplace=True,
|
| 85 |
+
)
|
| 86 |
+
|
| 87 |
+
ctx.save_for_backward(cos, sin, cu_seqlens)
|
| 88 |
+
ctx.max_seqlen = max_seqlen
|
| 89 |
+
return qkv
|
| 90 |
+
|
| 91 |
+
@staticmethod
|
| 92 |
+
def backward(ctx, do):
|
| 93 |
+
cos, sin, cu_seqlens = ctx.saved_tensors
|
| 94 |
+
do = do.contiguous()
|
| 95 |
+
total_nnz, _three, _nheads, headdim = do.shape
|
| 96 |
+
# We need dqkv to be contiguous so that when we reshape to combine (3, nheads) dimensions,
|
| 97 |
+
# we get the same tensor
|
| 98 |
+
dqk = do[:, :2].view(total_nnz, -1, headdim)
|
| 99 |
+
apply_rotary(
|
| 100 |
+
dqk,
|
| 101 |
+
cos,
|
| 102 |
+
sin,
|
| 103 |
+
seqlen_offsets=0,
|
| 104 |
+
cu_seqlens=cu_seqlens,
|
| 105 |
+
max_seqlen=ctx.max_seqlen,
|
| 106 |
+
interleaved=False,
|
| 107 |
+
inplace=True,
|
| 108 |
+
conjugate=True,
|
| 109 |
+
)
|
| 110 |
+
|
| 111 |
+
return do, None, None, None, None, None, None
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def apply_rotary_unpadded(
|
| 115 |
+
qkv,
|
| 116 |
+
cos,
|
| 117 |
+
sin,
|
| 118 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 119 |
+
max_seqlen: Optional[int] = None,
|
| 120 |
+
):
|
| 121 |
+
"""
|
| 122 |
+
Arguments:
|
| 123 |
+
qkv: (total_nnz, 3, nheads, headdim) - input tensor for packed QKV.
|
| 124 |
+
cos, sin: (seqlen_rotary, rotary_dim / 2)
|
| 125 |
+
interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
|
| 126 |
+
of 1st half and 2nd half (GPT-NeoX style).
|
| 127 |
+
inplace: if True, apply rotary embedding in-place.
|
| 128 |
+
seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
|
| 129 |
+
Most commonly used in inference when we have KV cache.
|
| 130 |
+
cu_seqlens: (batch + 1,) or None
|
| 131 |
+
max_seqlen: int
|
| 132 |
+
Return:
|
| 133 |
+
out: (total_nnz, dim)
|
| 134 |
+
rotary_dim must be <= headdim
|
| 135 |
+
Apply rotary embedding to the first rotary_dim of x.
|
| 136 |
+
"""
|
| 137 |
+
return ApplyRotaryEmbUnpad.apply(qkv, cos, sin, cu_seqlens, max_seqlen)
|
| 138 |
+
|
| 139 |
+
|
| 140 |
+
class ModernBertUnpaddedRotaryEmbedding(RotaryEmbedding):
|
| 141 |
+
"""
|
| 142 |
+
The rotary position embeddings applied directly to unpadded sequences.
|
| 143 |
+
"""
|
| 144 |
+
|
| 145 |
+
def __init__(
|
| 146 |
+
self,
|
| 147 |
+
dim: int,
|
| 148 |
+
base: float = 10000.0,
|
| 149 |
+
max_seqlen: Optional[int] = None,
|
| 150 |
+
device: Optional[torch.device] = None,
|
| 151 |
+
dtype: Optional[torch.dtype] = None,
|
| 152 |
+
):
|
| 153 |
+
"""
|
| 154 |
+
max_seqlen: if max_seqlen, device, and dtype are provided, we precompute the cos_sin_cache
|
| 155 |
+
up to max_seqlen. If the max_seqlen, device, or dtype during training/inference differ,
|
| 156 |
+
the cos_sin_cache wll be recomputed during the forward pass.
|
| 157 |
+
"""
|
| 158 |
+
super().__init__(dim=dim, base=base, pos_idx_in_fp32=True, device=device, interleaved=False)
|
| 159 |
+
self.max_seqlen = max_seqlen
|
| 160 |
+
|
| 161 |
+
if max_seqlen is not None and device is not None and dtype is not None:
|
| 162 |
+
self._update_cos_sin_cache(max_seqlen, device=device, dtype=dtype)
|
| 163 |
+
|
| 164 |
+
def forward(
|
| 165 |
+
self,
|
| 166 |
+
qkv: torch.Tensor,
|
| 167 |
+
cu_seqlens: torch.Tensor,
|
| 168 |
+
max_seqlen: Optional[int] = None,
|
| 169 |
+
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 170 |
+
"""
|
| 171 |
+
Apply rotary embedding *inplace* to qkv.
|
| 172 |
+
qkv: (total_nnz, 3, nheads, headdim)
|
| 173 |
+
cu_seqlens: (batch + 1,) cumulative sequence lengths
|
| 174 |
+
max_seqlen: int max seq length in the batch
|
| 175 |
+
"""
|
| 176 |
+
if max_seqlen is not None:
|
| 177 |
+
self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)
|
| 178 |
+
|
| 179 |
+
qkv = apply_rotary_unpadded(
|
| 180 |
+
qkv,
|
| 181 |
+
self._cos_cached,
|
| 182 |
+
self._sin_cached,
|
| 183 |
+
cu_seqlens=cu_seqlens,
|
| 184 |
+
max_seqlen=max_seqlen,
|
| 185 |
+
)
|
| 186 |
+
|
| 187 |
+
return qkv
|
| 188 |
+
|
| 189 |
+
def extra_repr(self) -> str:
|
| 190 |
+
return f"dim={self.dim}, base={self.base}, scale_base={self.scale_base}"
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
class ModernBertEmbeddings(nn.Module):
|
| 194 |
+
"""
|
| 195 |
+
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
|
| 196 |
+
"""
|
| 197 |
+
|
| 198 |
+
def __init__(self, config: ModernBertConfig):
|
| 199 |
+
super().__init__()
|
| 200 |
+
self.config = config
|
| 201 |
+
self.tok_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
|
| 202 |
+
self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
|
| 203 |
+
self.drop = nn.Dropout(config.embedding_dropout)
|
| 204 |
+
|
| 205 |
+
@torch.compile(dynamic=True)
|
| 206 |
+
def compiled_embeddings(self, input_ids: torch.LongTensor) -> torch.Tensor:
|
| 207 |
+
return self.drop(self.norm(self.tok_embeddings(input_ids)))
|
| 208 |
+
|
| 209 |
+
def forward(
|
| 210 |
+
self, input_ids: torch.LongTensor = None, inputs_embeds: Optional[torch.Tensor] = None
|
| 211 |
+
) -> torch.Tensor:
|
| 212 |
+
if inputs_embeds is not None:
|
| 213 |
+
hidden_states = self.drop(self.norm(inputs_embeds))
|
| 214 |
+
else:
|
| 215 |
+
hidden_states = (
|
| 216 |
+
self.compiled_embeddings(input_ids)
|
| 217 |
+
if self.config.reference_compile
|
| 218 |
+
else self.drop(self.norm(self.tok_embeddings(input_ids)))
|
| 219 |
+
)
|
| 220 |
+
return hidden_states
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
class ModernBertMLP(nn.Module):
|
| 224 |
+
"""Applies the GLU at the end of each ModernBERT layer.
|
| 225 |
+
|
| 226 |
+
Compared to the default BERT architecture, this block replaces :class:`~transformers.model.bert.modeling_bert.BertIntermediate`
|
| 227 |
+
and :class:`~transformers.model.bert.modeling_bert.SelfOutput` with a single module that has similar functionality.
|
| 228 |
+
"""
|
| 229 |
+
|
| 230 |
+
def __init__(self, config: ModernBertConfig):
|
| 231 |
+
super().__init__()
|
| 232 |
+
self.config = config
|
| 233 |
+
self.Wi = nn.Linear(config.hidden_size, int(config.intermediate_size) * 2, bias=config.mlp_bias)
|
| 234 |
+
self.act = ACT2FN[config.hidden_activation]
|
| 235 |
+
self.drop = nn.Dropout(config.mlp_dropout)
|
| 236 |
+
self.Wo = nn.Linear(config.intermediate_size, config.hidden_size, bias=config.mlp_bias)
|
| 237 |
+
|
| 238 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 239 |
+
input, gate = self.Wi(hidden_states).chunk(2, dim=-1)
|
| 240 |
+
return self.Wo(self.drop(self.act(input) * gate))
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
class ModernBertRotaryEmbedding(nn.Module):
|
| 244 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 245 |
+
super().__init__()
|
| 246 |
+
|
| 247 |
+
self.dim = dim
|
| 248 |
+
self.max_position_embeddings = max_position_embeddings
|
| 249 |
+
self.base = base
|
| 250 |
+
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
|
| 251 |
+
self.register_buffer("inv_freq", tensor=inv_freq, persistent=False)
|
| 252 |
+
|
| 253 |
+
@torch.no_grad()
|
| 254 |
+
def forward(self, x, position_ids, seq_len=None):
|
| 255 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
| 256 |
+
self.inv_freq.to(x.device)
|
| 257 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
| 258 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 259 |
+
# Force float32 since bfloat16 loses precision on long contexts
|
| 260 |
+
# See https://github.com/huggingface/transformers/pull/29285
|
| 261 |
+
device_type = x.device.type
|
| 262 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
| 263 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
| 264 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 265 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 266 |
+
cos = emb.cos()
|
| 267 |
+
sin = emb.sin()
|
| 268 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 269 |
+
|
| 270 |
+
|
| 271 |
+
def rotate_half(x):
|
| 272 |
+
"""Rotates half the hidden dims of the input."""
|
| 273 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 274 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 275 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
| 279 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
| 280 |
+
|
| 281 |
+
Args:
|
| 282 |
+
q (`torch.Tensor`): The query tensor.
|
| 283 |
+
k (`torch.Tensor`): The key tensor.
|
| 284 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
| 285 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
| 286 |
+
position_ids (`torch.Tensor`, *optional*):
|
| 287 |
+
Deprecated and unused.
|
| 288 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
| 289 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
| 290 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
| 291 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
| 292 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
| 293 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
| 294 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
| 295 |
+
Returns:
|
| 296 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
| 297 |
+
"""
|
| 298 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
| 299 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
| 300 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 301 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 302 |
+
return q_embed, k_embed
|
| 303 |
+
|
| 304 |
+
|
| 305 |
+
def eager_attention_forward(
|
| 306 |
+
module: "ModernBertAttention",
|
| 307 |
+
qkv: torch.Tensor,
|
| 308 |
+
attention_mask: torch.Tensor,
|
| 309 |
+
sliding_window_mask: torch.Tensor,
|
| 310 |
+
position_ids: Optional[torch.LongTensor],
|
| 311 |
+
local_attention: Tuple[int, int],
|
| 312 |
+
bs: int,
|
| 313 |
+
dim: int,
|
| 314 |
+
output_attentions: Optional[bool] = False,
|
| 315 |
+
**_kwargs,
|
| 316 |
+
) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
|
| 317 |
+
# qkv: [batch_size, seqlen, 3, nheads, headdim]
|
| 318 |
+
cos, sin = module.rotary_emb(qkv, position_ids=position_ids)
|
| 319 |
+
query, key, value = qkv.transpose(3, 1).unbind(dim=2)
|
| 320 |
+
# query, key, value: [batch_size, heads, seq_len, head_dim]
|
| 321 |
+
query, key = apply_rotary_pos_emb(query, key, cos, sin)
|
| 322 |
+
|
| 323 |
+
scale = module.head_dim**-0.5
|
| 324 |
+
attn_weights = torch.matmul(query, key.transpose(2, 3)) * scale
|
| 325 |
+
|
| 326 |
+
if local_attention != (-1, -1):
|
| 327 |
+
attention_mask = sliding_window_mask
|
| 328 |
+
|
| 329 |
+
attn_weights = attn_weights + attention_mask
|
| 330 |
+
|
| 331 |
+
# upcast attention to fp32
|
| 332 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
| 333 |
+
attn_weights = nn.functional.dropout(attn_weights, p=module.attention_dropout, training=module.training)
|
| 334 |
+
attn_output = torch.matmul(attn_weights, value)
|
| 335 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 336 |
+
attn_output = attn_output.view(bs, -1, dim)
|
| 337 |
+
if output_attentions:
|
| 338 |
+
return (attn_output, attn_weights)
|
| 339 |
+
return (attn_output,)
|
| 340 |
+
|
| 341 |
+
|
| 342 |
+
def flash_attention_forward(
|
| 343 |
+
module: "ModernBertAttention",
|
| 344 |
+
qkv: torch.Tensor,
|
| 345 |
+
rotary_emb: ModernBertUnpaddedRotaryEmbedding,
|
| 346 |
+
cu_seqlens: torch.Tensor,
|
| 347 |
+
max_seqlen: int,
|
| 348 |
+
local_attention: Tuple[int, int],
|
| 349 |
+
bs: int,
|
| 350 |
+
dim: int,
|
| 351 |
+
target_dtype: torch.dtype = torch.bfloat16,
|
| 352 |
+
**_kwargs,
|
| 353 |
+
) -> Tuple[torch.Tensor]:
|
| 354 |
+
# (total_seqlen, 3, nheads, headdim)
|
| 355 |
+
qkv = rotary_emb(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen)
|
| 356 |
+
|
| 357 |
+
convert_dtype = qkv.dtype not in (torch.float16, torch.bfloat16)
|
| 358 |
+
if convert_dtype:
|
| 359 |
+
# FA2 implementation only supports fp16 and bf16. If FA2 is supported,
|
| 360 |
+
# bfloat16 must be supported as of FA2 2.5.7. (Turing GPUs not supported)
|
| 361 |
+
orig_dtype = qkv.dtype
|
| 362 |
+
qkv = qkv.to(target_dtype)
|
| 363 |
+
|
| 364 |
+
attn = flash_attn_varlen_qkvpacked_func(
|
| 365 |
+
qkv,
|
| 366 |
+
cu_seqlens=cu_seqlens,
|
| 367 |
+
max_seqlen=max_seqlen,
|
| 368 |
+
dropout_p=module.attention_dropout if module.training else 0.0,
|
| 369 |
+
deterministic=module.deterministic_flash_attn,
|
| 370 |
+
window_size=local_attention,
|
| 371 |
+
)
|
| 372 |
+
attn = attn.to(orig_dtype) # type: ignore
|
| 373 |
+
else:
|
| 374 |
+
attn = flash_attn_varlen_qkvpacked_func(
|
| 375 |
+
qkv,
|
| 376 |
+
cu_seqlens=cu_seqlens,
|
| 377 |
+
max_seqlen=max_seqlen,
|
| 378 |
+
dropout_p=module.attention_dropout if module.training else 0.0,
|
| 379 |
+
deterministic=module.deterministic_flash_attn,
|
| 380 |
+
window_size=local_attention,
|
| 381 |
+
)
|
| 382 |
+
return (attn.view(bs, dim),)
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
def sdpa_attention_forward(
|
| 386 |
+
module: "ModernBertAttention",
|
| 387 |
+
qkv: torch.Tensor,
|
| 388 |
+
attention_mask: torch.Tensor,
|
| 389 |
+
sliding_window_mask: torch.Tensor,
|
| 390 |
+
position_ids: Optional[torch.LongTensor],
|
| 391 |
+
local_attention: Tuple[int, int],
|
| 392 |
+
bs: int,
|
| 393 |
+
dim: int,
|
| 394 |
+
**_kwargs,
|
| 395 |
+
) -> Tuple[torch.Tensor]:
|
| 396 |
+
# qkv: [batch_size, seqlen, 3, nheads, headdim]
|
| 397 |
+
cos, sin = module.rotary_emb(qkv, position_ids=position_ids)
|
| 398 |
+
query, key, value = qkv.transpose(3, 1).unbind(dim=2)
|
| 399 |
+
# query, key, value: [batch_size, heads, seq_len, head_dim]
|
| 400 |
+
query, key = apply_rotary_pos_emb(query, key, cos, sin)
|
| 401 |
+
|
| 402 |
+
if local_attention != (-1, -1):
|
| 403 |
+
attention_mask = sliding_window_mask
|
| 404 |
+
|
| 405 |
+
attn_output = (
|
| 406 |
+
F.scaled_dot_product_attention(
|
| 407 |
+
query,
|
| 408 |
+
key,
|
| 409 |
+
value,
|
| 410 |
+
dropout_p=module.attention_dropout if module.training else 0.0,
|
| 411 |
+
attn_mask=attention_mask,
|
| 412 |
+
)
|
| 413 |
+
.transpose(1, 2)
|
| 414 |
+
.contiguous()
|
| 415 |
+
)
|
| 416 |
+
attn_output = attn_output.view(bs, -1, dim)
|
| 417 |
+
return (attn_output,)
|
| 418 |
+
|
| 419 |
+
|
| 420 |
+
MODERNBERT_ATTENTION_FUNCTION = {
|
| 421 |
+
"flash_attention_2": flash_attention_forward,
|
| 422 |
+
"eager": eager_attention_forward,
|
| 423 |
+
"sdpa": sdpa_attention_forward,
|
| 424 |
+
}
|
| 425 |
+
|
| 426 |
+
|
| 427 |
+
class ModernBertAttention(nn.Module):
|
| 428 |
+
"""Performs multi-headed self attention on a batch of unpadded sequences.
|
| 429 |
+
|
| 430 |
+
If Flash Attention 2 is installed, this module uses Flash Attention to improve throughput.
|
| 431 |
+
If Flash Attention 2 is not installed, the implementation will use PyTorch's SDPA kernel,
|
| 432 |
+
which requires padding and unpadding inputs, adding some overhead.
|
| 433 |
+
|
| 434 |
+
See `forward` method for additional details.
|
| 435 |
+
"""
|
| 436 |
+
|
| 437 |
+
def __init__(self, config: ModernBertConfig, layer_id: Optional[int] = None):
|
| 438 |
+
super().__init__()
|
| 439 |
+
self.config = config
|
| 440 |
+
self.layer_id = layer_id
|
| 441 |
+
|
| 442 |
+
if config.hidden_size % config.num_attention_heads != 0:
|
| 443 |
+
raise ValueError(
|
| 444 |
+
f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention heads ({config.num_attention_heads})"
|
| 445 |
+
)
|
| 446 |
+
|
| 447 |
+
self.attention_dropout = config.attention_dropout
|
| 448 |
+
self.deterministic_flash_attn = config.deterministic_flash_attn
|
| 449 |
+
self.num_heads = config.num_attention_heads
|
| 450 |
+
self.head_dim = config.hidden_size // config.num_attention_heads
|
| 451 |
+
self.all_head_size = self.head_dim * self.num_heads
|
| 452 |
+
self.Wqkv = nn.Linear(config.hidden_size, 3 * self.all_head_size, bias=config.attention_bias)
|
| 453 |
+
|
| 454 |
+
if layer_id % config.global_attn_every_n_layers != 0:
|
| 455 |
+
self.local_attention = (config.local_attention // 2, config.local_attention // 2)
|
| 456 |
+
else:
|
| 457 |
+
self.local_attention = (-1, -1)
|
| 458 |
+
|
| 459 |
+
rope_theta = config.global_rope_theta
|
| 460 |
+
max_position_embeddings = config.max_position_embeddings
|
| 461 |
+
if self.local_attention != (-1, -1):
|
| 462 |
+
if config.local_rope_theta is not None:
|
| 463 |
+
rope_theta = config.local_rope_theta
|
| 464 |
+
max_position_embeddings = config.local_attention
|
| 465 |
+
|
| 466 |
+
if config._attn_implementation == "flash_attention_2":
|
| 467 |
+
self.rotary_emb = ModernBertUnpaddedRotaryEmbedding(
|
| 468 |
+
dim=self.head_dim, max_seqlen=max_position_embeddings, base=rope_theta
|
| 469 |
+
)
|
| 470 |
+
else:
|
| 471 |
+
self.rotary_emb = ModernBertRotaryEmbedding(
|
| 472 |
+
dim=self.head_dim, max_position_embeddings=max_position_embeddings, base=rope_theta
|
| 473 |
+
)
|
| 474 |
+
|
| 475 |
+
self.Wo = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias)
|
| 476 |
+
self.out_drop = nn.Dropout(config.attention_dropout) if config.attention_dropout > 0.0 else nn.Identity()
|
| 477 |
+
self.pruned_heads = set()
|
| 478 |
+
|
| 479 |
+
def forward(
|
| 480 |
+
self,
|
| 481 |
+
hidden_states: torch.Tensor,
|
| 482 |
+
output_attentions: Optional[bool] = False,
|
| 483 |
+
**kwargs,
|
| 484 |
+
) -> torch.Tensor:
|
| 485 |
+
qkv = self.Wqkv(hidden_states)
|
| 486 |
+
|
| 487 |
+
bs = hidden_states.shape[0]
|
| 488 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 489 |
+
qkv = qkv.view(-1, 3, self.num_heads, self.head_dim)
|
| 490 |
+
else:
|
| 491 |
+
qkv = qkv.view(bs, -1, 3, self.num_heads, self.head_dim)
|
| 492 |
+
|
| 493 |
+
attn_outputs = MODERNBERT_ATTENTION_FUNCTION[self.config._attn_implementation](
|
| 494 |
+
self,
|
| 495 |
+
qkv=qkv,
|
| 496 |
+
rotary_emb=self.rotary_emb,
|
| 497 |
+
local_attention=self.local_attention,
|
| 498 |
+
bs=bs,
|
| 499 |
+
dim=self.all_head_size,
|
| 500 |
+
output_attentions=output_attentions,
|
| 501 |
+
**kwargs,
|
| 502 |
+
)
|
| 503 |
+
hidden_states = attn_outputs[0]
|
| 504 |
+
hidden_states = self.out_drop(self.Wo(hidden_states))
|
| 505 |
+
|
| 506 |
+
return (hidden_states,) + attn_outputs[1:] # add attentions if outputted
|
| 507 |
+
|
| 508 |
+
|
| 509 |
+
class ModernBertEncoderLayer(nn.Module):
|
| 510 |
+
def __init__(self, config: ModernBertConfig, layer_id: Optional[int] = None):
|
| 511 |
+
super().__init__()
|
| 512 |
+
self.config = config
|
| 513 |
+
if layer_id == 0:
|
| 514 |
+
self.attn_norm = nn.Identity()
|
| 515 |
+
else:
|
| 516 |
+
self.attn_norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
|
| 517 |
+
self.attn = ModernBertAttention(config=config, layer_id=layer_id)
|
| 518 |
+
self.mlp_norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
|
| 519 |
+
self.mlp = ModernBertMLP(config)
|
| 520 |
+
|
| 521 |
+
@torch.compile(dynamic=True)
|
| 522 |
+
def compiled_mlp(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 523 |
+
return self.mlp(self.mlp_norm(hidden_states))
|
| 524 |
+
|
| 525 |
+
def forward(
|
| 526 |
+
self,
|
| 527 |
+
hidden_states: torch.Tensor,
|
| 528 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 529 |
+
sliding_window_mask: Optional[torch.Tensor] = None,
|
| 530 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 531 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 532 |
+
max_seqlen: Optional[int] = None,
|
| 533 |
+
output_attentions: Optional[bool] = False,
|
| 534 |
+
) -> torch.Tensor:
|
| 535 |
+
attn_outputs = self.attn(
|
| 536 |
+
self.attn_norm(hidden_states),
|
| 537 |
+
attention_mask=attention_mask,
|
| 538 |
+
sliding_window_mask=sliding_window_mask,
|
| 539 |
+
position_ids=position_ids,
|
| 540 |
+
cu_seqlens=cu_seqlens,
|
| 541 |
+
max_seqlen=max_seqlen,
|
| 542 |
+
output_attentions=output_attentions,
|
| 543 |
+
)
|
| 544 |
+
hidden_states = hidden_states + attn_outputs[0]
|
| 545 |
+
mlp_output = (
|
| 546 |
+
self.compiled_mlp(hidden_states)
|
| 547 |
+
if self.config.reference_compile
|
| 548 |
+
else self.mlp(self.mlp_norm(hidden_states))
|
| 549 |
+
)
|
| 550 |
+
hidden_states = hidden_states + mlp_output
|
| 551 |
+
|
| 552 |
+
return (hidden_states,) + attn_outputs[1:] # add attentions if outputted
|
| 553 |
+
|
| 554 |
+
|
| 555 |
+
MODERNBERT_START_DOCSTRING = r"""
|
| 556 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 557 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 558 |
+
etc.)
|
| 559 |
+
|
| 560 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 561 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 562 |
+
and behavior.
|
| 563 |
+
|
| 564 |
+
Parameters:
|
| 565 |
+
config ([`ModernBertConfig`]):
|
| 566 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
| 567 |
+
load the weights associated with the model, only the configuration. Check out the
|
| 568 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 569 |
+
"""
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
@add_start_docstrings(
|
| 573 |
+
"The bare ModernBert Model outputting raw hidden-states without any specific head on top.",
|
| 574 |
+
MODERNBERT_START_DOCSTRING,
|
| 575 |
+
)
|
| 576 |
+
class ModernBertPreTrainedModel(PreTrainedModel):
|
| 577 |
+
config_class = ModernBertConfig
|
| 578 |
+
base_model_prefix = "model"
|
| 579 |
+
supports_gradient_checkpointing = True
|
| 580 |
+
_no_split_modules = ["ModernBertEmbeddings", "ModernBertEncoderLayer"]
|
| 581 |
+
_supports_flash_attn_2 = True
|
| 582 |
+
_supports_sdpa = True
|
| 583 |
+
_supports_flex_attn = False
|
| 584 |
+
|
| 585 |
+
def _init_weights(self, module: nn.Module):
|
| 586 |
+
cutoff_factor = self.config.initializer_cutoff_factor
|
| 587 |
+
if cutoff_factor is None:
|
| 588 |
+
cutoff_factor = 3
|
| 589 |
+
|
| 590 |
+
def init_weight(module: nn.Module, std: float):
|
| 591 |
+
nn.init.trunc_normal_(
|
| 592 |
+
module.weight,
|
| 593 |
+
mean=0.0,
|
| 594 |
+
std=std,
|
| 595 |
+
a=-cutoff_factor * std,
|
| 596 |
+
b=cutoff_factor * std,
|
| 597 |
+
)
|
| 598 |
+
|
| 599 |
+
if isinstance(module, nn.Linear):
|
| 600 |
+
if module.bias is not None:
|
| 601 |
+
nn.init.zeros_(module.bias)
|
| 602 |
+
|
| 603 |
+
stds = {
|
| 604 |
+
"in": self.config.initializer_range,
|
| 605 |
+
"out": self.config.initializer_range / math.sqrt(2.0 * self.config.num_hidden_layers),
|
| 606 |
+
"embedding": self.config.initializer_range,
|
| 607 |
+
"final_out": self.config.hidden_size**-0.5,
|
| 608 |
+
}
|
| 609 |
+
|
| 610 |
+
if isinstance(module, ModernBertEmbeddings):
|
| 611 |
+
init_weight(module.tok_embeddings, stds["embedding"])
|
| 612 |
+
elif isinstance(module, ModernBertMLP):
|
| 613 |
+
init_weight(module.Wi, stds["in"])
|
| 614 |
+
init_weight(module.Wo, stds["out"])
|
| 615 |
+
elif isinstance(module, ModernBertAttention):
|
| 616 |
+
init_weight(module.Wqkv, stds["in"])
|
| 617 |
+
init_weight(module.Wo, stds["out"])
|
| 618 |
+
elif isinstance(module, ModernBertPredictionHead):
|
| 619 |
+
init_weight(module.dense, stds["out"])
|
| 620 |
+
elif isinstance(module, ModernBertForMaskedLM):
|
| 621 |
+
init_weight(module.decoder, stds["out"])
|
| 622 |
+
elif isinstance(module, (ModernBertForSequenceClassification, ModernBertForTokenClassification)):
|
| 623 |
+
init_weight(module.classifier, stds["final_out"])
|
| 624 |
+
|
| 625 |
+
@classmethod
|
| 626 |
+
def _autoset_attn_implementation(
|
| 627 |
+
cls,
|
| 628 |
+
config,
|
| 629 |
+
use_flash_attention_2: bool = False,
|
| 630 |
+
torch_dtype: Optional[torch.dtype] = None,
|
| 631 |
+
device_map: Optional[Union[str, Dict[str, int]]] = None,
|
| 632 |
+
check_device_map: bool = True,
|
| 633 |
+
):
|
| 634 |
+
# If the user didn't specify anything, try to use flash_attention_2 if available.
|
| 635 |
+
# Otherwise we fall back to the default SDPA -> Eager from the super() method.
|
| 636 |
+
if config._attn_implementation_internal is None:
|
| 637 |
+
config._attn_implementation_internal = "flash_attention_2"
|
| 638 |
+
try:
|
| 639 |
+
return cls._check_and_enable_flash_attn_2(
|
| 640 |
+
config,
|
| 641 |
+
torch_dtype=torch_dtype,
|
| 642 |
+
device_map=device_map,
|
| 643 |
+
hard_check_only=False,
|
| 644 |
+
check_device_map=check_device_map,
|
| 645 |
+
)
|
| 646 |
+
except (ValueError, ImportError):
|
| 647 |
+
config._attn_implementation_internal = None
|
| 648 |
+
return super()._autoset_attn_implementation(
|
| 649 |
+
config,
|
| 650 |
+
use_flash_attention_2=use_flash_attention_2,
|
| 651 |
+
torch_dtype=torch_dtype,
|
| 652 |
+
device_map=device_map,
|
| 653 |
+
check_device_map=check_device_map,
|
| 654 |
+
)
|
| 655 |
+
|
| 656 |
+
def _maybe_set_compile(self):
|
| 657 |
+
if self.config.reference_compile is False:
|
| 658 |
+
return
|
| 659 |
+
|
| 660 |
+
if hasattr(self, "hf_device_map") and len(self.hf_device_map) > 1:
|
| 661 |
+
if self.config.reference_compile:
|
| 662 |
+
logger.warning_once(
|
| 663 |
+
"If `accelerate` split the model across devices, `torch.compile` will not work. "
|
| 664 |
+
"Falling back to non-compiled mode."
|
| 665 |
+
)
|
| 666 |
+
self.config.reference_compile = False
|
| 667 |
+
|
| 668 |
+
if self.device.type == "mps":
|
| 669 |
+
if self.config.reference_compile:
|
| 670 |
+
logger.warning_once(
|
| 671 |
+
"Compiling the model with `torch.compile` and using a `torch.mps` device is not supported. "
|
| 672 |
+
"Falling back to non-compiled mode."
|
| 673 |
+
)
|
| 674 |
+
self.config.reference_compile = False
|
| 675 |
+
|
| 676 |
+
if self.config.reference_compile is None:
|
| 677 |
+
self.config.reference_compile = is_triton_available()
|
| 678 |
+
|
| 679 |
+
def resize_token_embeddings(self, *args, **kwargs):
|
| 680 |
+
model_embeds = super().resize_token_embeddings(*args, **kwargs)
|
| 681 |
+
|
| 682 |
+
if self.config.reference_compile in {True, None}:
|
| 683 |
+
if self.config.reference_compile:
|
| 684 |
+
logger.warning_once(
|
| 685 |
+
"Resizing token embeddings with `torch.compile` is not supported. Falling back to non-compiled mode."
|
| 686 |
+
)
|
| 687 |
+
self.config.reference_compile = False
|
| 688 |
+
|
| 689 |
+
return model_embeds
|
| 690 |
+
|
| 691 |
+
|
| 692 |
+
def _unpad_modernbert_input(
|
| 693 |
+
inputs: torch.Tensor,
|
| 694 |
+
attention_mask: torch.Tensor,
|
| 695 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 696 |
+
labels: Optional[torch.Tensor] = None,
|
| 697 |
+
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, Optional[torch.Tensor], Optional[torch.Tensor]]:
|
| 698 |
+
"""
|
| 699 |
+
Remove padding from input sequences.
|
| 700 |
+
|
| 701 |
+
Args:
|
| 702 |
+
inputs: (batch, seqlen, ...) or (batch, seqlen)
|
| 703 |
+
attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
|
| 704 |
+
position_ids: (batch, seqlen), int, position ids
|
| 705 |
+
labels: (batch, seqlen), int, labels
|
| 706 |
+
|
| 707 |
+
Returns:
|
| 708 |
+
unpadded_inputs: (total_nnz, ...), where total_nnz = number of tokens selected in attention_mask.
|
| 709 |
+
indices: (total_nnz)
|
| 710 |
+
cu_seqlens: (batch + 1), the cumulative sequence lengths
|
| 711 |
+
max_seqlen_in_batch: int
|
| 712 |
+
unpadded_position_ids: (total_nnz) or None
|
| 713 |
+
unpadded_labels: (total_nnz) or None
|
| 714 |
+
"""
|
| 715 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
| 716 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
| 717 |
+
max_seqlen_in_batch = int(seqlens_in_batch.max().item())
|
| 718 |
+
cu_seqlens = torch.nn.functional.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
|
| 719 |
+
|
| 720 |
+
if inputs.dim() == 2:
|
| 721 |
+
unpadded_inputs = inputs.flatten()[indices]
|
| 722 |
+
else:
|
| 723 |
+
batch, seqlen, *rest = inputs.shape
|
| 724 |
+
shape = batch * seqlen
|
| 725 |
+
unpadded_inputs = inputs.view(shape, *rest)[indices]
|
| 726 |
+
|
| 727 |
+
unpadded_position_ids = position_ids.flatten()[indices] if position_ids is not None else None
|
| 728 |
+
unpadded_labels = labels.flatten()[indices] if labels is not None else None
|
| 729 |
+
|
| 730 |
+
return unpadded_inputs, indices, cu_seqlens, max_seqlen_in_batch, unpadded_position_ids, unpadded_labels
|
| 731 |
+
|
| 732 |
+
|
| 733 |
+
def _pad_modernbert_output(
|
| 734 |
+
inputs: torch.Tensor,
|
| 735 |
+
indices: torch.Tensor,
|
| 736 |
+
batch: int,
|
| 737 |
+
seqlen: int,
|
| 738 |
+
) -> torch.Tensor:
|
| 739 |
+
"""
|
| 740 |
+
Add padding to sequences.
|
| 741 |
+
|
| 742 |
+
Args:
|
| 743 |
+
inputs: (total_nnz, ...) or (total_nnz,), where total_nnz = number of tokens selected in attention_mask.
|
| 744 |
+
indices: (total_nnz)
|
| 745 |
+
batch: int, batch size
|
| 746 |
+
seqlen: int, max sequence length
|
| 747 |
+
|
| 748 |
+
Returns:
|
| 749 |
+
padded_inputs: (batch, seqlen, ...) or (batch, seqlen)
|
| 750 |
+
"""
|
| 751 |
+
if inputs.dim() == 1:
|
| 752 |
+
output = torch.zeros(batch * seqlen, dtype=inputs.dtype, device=inputs.device)
|
| 753 |
+
output[indices] = inputs
|
| 754 |
+
padded_inputs = output.view(batch, seqlen)
|
| 755 |
+
else:
|
| 756 |
+
_, *rest = inputs.shape
|
| 757 |
+
output = torch.zeros(batch * seqlen, *rest, dtype=inputs.dtype, device=inputs.device)
|
| 758 |
+
output[indices] = inputs
|
| 759 |
+
padded_inputs = output.view(batch, seqlen, *rest)
|
| 760 |
+
|
| 761 |
+
return padded_inputs
|
| 762 |
+
|
| 763 |
+
|
| 764 |
+
MODERNBERT_INPUTS_DOCSTRING = r"""
|
| 765 |
+
Args:
|
| 766 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 767 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 768 |
+
it.
|
| 769 |
+
|
| 770 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 771 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 772 |
+
|
| 773 |
+
[What are input IDs?](../glossary#input-ids)
|
| 774 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 775 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 776 |
+
|
| 777 |
+
- 1 for tokens that are **not masked**,
|
| 778 |
+
- 0 for tokens that are **masked**.
|
| 779 |
+
|
| 780 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 781 |
+
|
| 782 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 783 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 784 |
+
|
| 785 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
| 786 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
| 787 |
+
information on the default strategy.
|
| 788 |
+
|
| 789 |
+
- 1 indicates the head is **not masked**,
|
| 790 |
+
- 0 indicates the head is **masked**.
|
| 791 |
+
sliding_window_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 792 |
+
Mask to avoid performing attention on padding or far-away tokens. In ModernBert, only every few layers
|
| 793 |
+
perform global attention, while the rest perform local attention. This mask is used to avoid attending to
|
| 794 |
+
far-away tokens in the local attention layers.
|
| 795 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 796 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 797 |
+
config.n_positions - 1]`.
|
| 798 |
+
|
| 799 |
+
[What are position IDs?](../glossary#position-ids)
|
| 800 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
| 801 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
| 802 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
| 803 |
+
model's internal embedding lookup matrix.
|
| 804 |
+
indices (`torch.Tensor` of shape `(total_unpadded_tokens,)`, *optional*):
|
| 805 |
+
Indices of the non-padding tokens in the input sequence. Used for unpadding the output.
|
| 806 |
+
cu_seqlens (`torch.Tensor` of shape `(batch + 1,)`, *optional*):
|
| 807 |
+
Cumulative sequence lengths of the input sequences. Used to index the unpadded tensors.
|
| 808 |
+
max_seqlen (`int`, *optional*):
|
| 809 |
+
Maximum sequence length in the batch. Used to pad the output tensors.
|
| 810 |
+
batch_size (`int`, *optional*):
|
| 811 |
+
Batch size of the input sequences. Used to pad the output tensors.
|
| 812 |
+
seq_len (`int`, *optional*):
|
| 813 |
+
Sequence length of the input sequences. Used to pad the output tensors.
|
| 814 |
+
output_attentions (`bool`, *optional*):
|
| 815 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 816 |
+
tensors for more detail.
|
| 817 |
+
output_hidden_states (`bool`, *optional*):
|
| 818 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 819 |
+
more detail.
|
| 820 |
+
return_dict (`bool`, *optional*):
|
| 821 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 822 |
+
"""
|
| 823 |
+
|
| 824 |
+
|
| 825 |
+
@add_start_docstrings(
|
| 826 |
+
"The bare ModernBert Model outputting raw hidden-states without any specific head on top.",
|
| 827 |
+
MODERNBERT_START_DOCSTRING,
|
| 828 |
+
)
|
| 829 |
+
class ModernBertModel(ModernBertPreTrainedModel):
|
| 830 |
+
def __init__(self, config: ModernBertConfig):
|
| 831 |
+
super().__init__(config)
|
| 832 |
+
self.config = config
|
| 833 |
+
self.embeddings = ModernBertEmbeddings(config)
|
| 834 |
+
self.layers = nn.ModuleList(
|
| 835 |
+
[ModernBertEncoderLayer(config, layer_id) for layer_id in range(config.num_hidden_layers)]
|
| 836 |
+
)
|
| 837 |
+
self.final_norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
|
| 838 |
+
self.gradient_checkpointing = False
|
| 839 |
+
self.post_init()
|
| 840 |
+
|
| 841 |
+
def get_input_embeddings(self):
|
| 842 |
+
return self.embeddings.tok_embeddings
|
| 843 |
+
|
| 844 |
+
def set_input_embeddings(self, value):
|
| 845 |
+
self.embeddings.tok_embeddings = value
|
| 846 |
+
|
| 847 |
+
@add_start_docstrings_to_model_forward(MODERNBERT_INPUTS_DOCSTRING)
|
| 848 |
+
@add_code_sample_docstrings(
|
| 849 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 850 |
+
output_type=BaseModelOutput,
|
| 851 |
+
config_class=_CONFIG_FOR_DOC,
|
| 852 |
+
)
|
| 853 |
+
def forward(
|
| 854 |
+
self,
|
| 855 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 856 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 857 |
+
sliding_window_mask: Optional[torch.Tensor] = None,
|
| 858 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 859 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 860 |
+
indices: Optional[torch.Tensor] = None,
|
| 861 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 862 |
+
max_seqlen: Optional[int] = None,
|
| 863 |
+
batch_size: Optional[int] = None,
|
| 864 |
+
seq_len: Optional[int] = None,
|
| 865 |
+
output_attentions: Optional[bool] = None,
|
| 866 |
+
output_hidden_states: Optional[bool] = None,
|
| 867 |
+
return_dict: Optional[bool] = None,
|
| 868 |
+
) -> Union[Tuple[torch.Tensor, ...], BaseModelOutput]:
|
| 869 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 870 |
+
output_hidden_states = (
|
| 871 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 872 |
+
)
|
| 873 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 874 |
+
|
| 875 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
| 876 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
| 877 |
+
|
| 878 |
+
all_hidden_states = () if output_hidden_states else None
|
| 879 |
+
all_self_attentions = () if output_attentions else None
|
| 880 |
+
|
| 881 |
+
self._maybe_set_compile()
|
| 882 |
+
|
| 883 |
+
if input_ids is not None:
|
| 884 |
+
self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
|
| 885 |
+
|
| 886 |
+
if batch_size is None and seq_len is None:
|
| 887 |
+
if inputs_embeds is not None:
|
| 888 |
+
batch_size, seq_len = inputs_embeds.shape[:2]
|
| 889 |
+
else:
|
| 890 |
+
batch_size, seq_len = input_ids.shape[:2]
|
| 891 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
| 892 |
+
|
| 893 |
+
if attention_mask is None:
|
| 894 |
+
attention_mask = torch.ones((batch_size, seq_len), device=device, dtype=torch.bool)
|
| 895 |
+
|
| 896 |
+
repad = False
|
| 897 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 898 |
+
if indices is None and cu_seqlens is None and max_seqlen is None:
|
| 899 |
+
repad = True
|
| 900 |
+
if inputs_embeds is None:
|
| 901 |
+
with torch.no_grad():
|
| 902 |
+
input_ids, indices, cu_seqlens, max_seqlen, *_ = _unpad_modernbert_input(
|
| 903 |
+
inputs=input_ids, attention_mask=attention_mask
|
| 904 |
+
)
|
| 905 |
+
else:
|
| 906 |
+
inputs_embeds, indices, cu_seqlens, max_seqlen, *_ = _unpad_modernbert_input(
|
| 907 |
+
inputs=inputs_embeds, attention_mask=attention_mask
|
| 908 |
+
)
|
| 909 |
+
else:
|
| 910 |
+
if position_ids is None:
|
| 911 |
+
position_ids = torch.arange(seq_len, device=device).unsqueeze(0)
|
| 912 |
+
|
| 913 |
+
attention_mask, sliding_window_mask = self._update_attention_mask(
|
| 914 |
+
attention_mask, output_attentions=output_attentions
|
| 915 |
+
)
|
| 916 |
+
|
| 917 |
+
hidden_states = self.embeddings(input_ids=input_ids, inputs_embeds=inputs_embeds)
|
| 918 |
+
|
| 919 |
+
for encoder_layer in self.layers:
|
| 920 |
+
if output_hidden_states:
|
| 921 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 922 |
+
|
| 923 |
+
if self.gradient_checkpointing and self.training:
|
| 924 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 925 |
+
encoder_layer.__call__,
|
| 926 |
+
hidden_states,
|
| 927 |
+
attention_mask,
|
| 928 |
+
sliding_window_mask,
|
| 929 |
+
position_ids,
|
| 930 |
+
cu_seqlens,
|
| 931 |
+
max_seqlen,
|
| 932 |
+
output_attentions,
|
| 933 |
+
)
|
| 934 |
+
else:
|
| 935 |
+
layer_outputs = encoder_layer(
|
| 936 |
+
hidden_states,
|
| 937 |
+
attention_mask=attention_mask,
|
| 938 |
+
sliding_window_mask=sliding_window_mask,
|
| 939 |
+
position_ids=position_ids,
|
| 940 |
+
cu_seqlens=cu_seqlens,
|
| 941 |
+
max_seqlen=max_seqlen,
|
| 942 |
+
output_attentions=output_attentions,
|
| 943 |
+
)
|
| 944 |
+
hidden_states = layer_outputs[0]
|
| 945 |
+
if output_attentions and len(layer_outputs) > 1:
|
| 946 |
+
all_self_attentions = all_self_attentions + (layer_outputs[1],)
|
| 947 |
+
|
| 948 |
+
if output_hidden_states:
|
| 949 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 950 |
+
|
| 951 |
+
hidden_states = self.final_norm(hidden_states)
|
| 952 |
+
|
| 953 |
+
if repad:
|
| 954 |
+
hidden_states = _pad_modernbert_output(
|
| 955 |
+
inputs=hidden_states, indices=indices, batch=batch_size, seqlen=seq_len
|
| 956 |
+
)
|
| 957 |
+
if all_hidden_states is not None:
|
| 958 |
+
all_hidden_states = tuple(
|
| 959 |
+
_pad_modernbert_output(inputs=hs, indices=indices, batch=batch_size, seqlen=seq_len)
|
| 960 |
+
for hs in all_hidden_states
|
| 961 |
+
)
|
| 962 |
+
|
| 963 |
+
if not return_dict:
|
| 964 |
+
return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
|
| 965 |
+
return BaseModelOutput(
|
| 966 |
+
last_hidden_state=hidden_states,
|
| 967 |
+
hidden_states=all_hidden_states,
|
| 968 |
+
attentions=all_self_attentions,
|
| 969 |
+
)
|
| 970 |
+
|
| 971 |
+
def _update_attention_mask(self, attention_mask: torch.Tensor, output_attentions: bool) -> torch.Tensor:
|
| 972 |
+
if output_attentions:
|
| 973 |
+
if self.config._attn_implementation == "sdpa":
|
| 974 |
+
logger.warning_once(
|
| 975 |
+
"Outputting attentions is only supported with the 'eager' attention implementation, "
|
| 976 |
+
'not with "sdpa". Falling back to `attn_implementation="eager"`.'
|
| 977 |
+
)
|
| 978 |
+
self.config._attn_implementation = "eager"
|
| 979 |
+
elif self.config._attn_implementation != "eager":
|
| 980 |
+
logger.warning_once(
|
| 981 |
+
"Outputting attentions is only supported with the eager attention implementation, "
|
| 982 |
+
f'not with {self.config._attn_implementation}. Consider setting `attn_implementation="eager"`.'
|
| 983 |
+
" Setting `output_attentions=False`."
|
| 984 |
+
)
|
| 985 |
+
|
| 986 |
+
global_attention_mask = _prepare_4d_attention_mask(attention_mask, self.dtype)
|
| 987 |
+
|
| 988 |
+
# Create position indices
|
| 989 |
+
rows = torch.arange(global_attention_mask.shape[2]).unsqueeze(0)
|
| 990 |
+
# Calculate distance between positions
|
| 991 |
+
distance = torch.abs(rows - rows.T)
|
| 992 |
+
|
| 993 |
+
# Create sliding window mask (1 for positions within window, 0 outside)
|
| 994 |
+
window_mask = (
|
| 995 |
+
(distance <= self.config.local_attention // 2).unsqueeze(0).unsqueeze(0).to(attention_mask.device)
|
| 996 |
+
)
|
| 997 |
+
# Combine with existing mask
|
| 998 |
+
sliding_window_mask = global_attention_mask.masked_fill(window_mask.logical_not(), torch.finfo(self.dtype).min)
|
| 999 |
+
|
| 1000 |
+
return global_attention_mask, sliding_window_mask
|
| 1001 |
+
|
| 1002 |
+
|
| 1003 |
+
class ModernBertPredictionHead(nn.Module):
|
| 1004 |
+
def __init__(self, config: ModernBertConfig):
|
| 1005 |
+
super().__init__()
|
| 1006 |
+
self.config = config
|
| 1007 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size, config.classifier_bias)
|
| 1008 |
+
self.act = ACT2FN[config.classifier_activation]
|
| 1009 |
+
self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
|
| 1010 |
+
|
| 1011 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 1012 |
+
return self.norm(self.act(self.dense(hidden_states)))
|
| 1013 |
+
|
| 1014 |
+
|
| 1015 |
+
@add_start_docstrings(
|
| 1016 |
+
"The ModernBert Model with a decoder head on top that is used for masked language modeling.",
|
| 1017 |
+
MODERNBERT_START_DOCSTRING,
|
| 1018 |
+
)
|
| 1019 |
+
class ModernBertForMaskedLM(ModernBertPreTrainedModel):
|
| 1020 |
+
_tied_weights_keys = ["decoder.weight"]
|
| 1021 |
+
|
| 1022 |
+
def __init__(self, config: ModernBertConfig):
|
| 1023 |
+
super().__init__(config)
|
| 1024 |
+
self.config = config
|
| 1025 |
+
self.model = ModernBertModel(config)
|
| 1026 |
+
self.head = ModernBertPredictionHead(config)
|
| 1027 |
+
self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=config.decoder_bias)
|
| 1028 |
+
|
| 1029 |
+
self.sparse_prediction = self.config.sparse_prediction
|
| 1030 |
+
self.sparse_pred_ignore_index = self.config.sparse_pred_ignore_index
|
| 1031 |
+
|
| 1032 |
+
# Initialize weights and apply final processing
|
| 1033 |
+
self.post_init()
|
| 1034 |
+
|
| 1035 |
+
def get_output_embeddings(self):
|
| 1036 |
+
return self.decoder
|
| 1037 |
+
|
| 1038 |
+
def set_output_embeddings(self, new_embeddings: nn.Linear):
|
| 1039 |
+
self.decoder = new_embeddings
|
| 1040 |
+
|
| 1041 |
+
@torch.compile(dynamic=True)
|
| 1042 |
+
def compiled_head(self, output: torch.Tensor) -> torch.Tensor:
|
| 1043 |
+
return self.decoder(self.head(output))
|
| 1044 |
+
|
| 1045 |
+
@add_start_docstrings_to_model_forward(MODERNBERT_INPUTS_DOCSTRING)
|
| 1046 |
+
@add_code_sample_docstrings(
|
| 1047 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1048 |
+
output_type=MaskedLMOutput,
|
| 1049 |
+
config_class=_CONFIG_FOR_DOC,
|
| 1050 |
+
)
|
| 1051 |
+
def forward(
|
| 1052 |
+
self,
|
| 1053 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1054 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1055 |
+
sliding_window_mask: Optional[torch.Tensor] = None,
|
| 1056 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1057 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1058 |
+
labels: Optional[torch.Tensor] = None,
|
| 1059 |
+
indices: Optional[torch.Tensor] = None,
|
| 1060 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 1061 |
+
max_seqlen: Optional[int] = None,
|
| 1062 |
+
batch_size: Optional[int] = None,
|
| 1063 |
+
seq_len: Optional[int] = None,
|
| 1064 |
+
output_attentions: Optional[bool] = None,
|
| 1065 |
+
output_hidden_states: Optional[bool] = None,
|
| 1066 |
+
return_dict: Optional[bool] = None,
|
| 1067 |
+
**kwargs,
|
| 1068 |
+
) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
|
| 1069 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1070 |
+
self._maybe_set_compile()
|
| 1071 |
+
|
| 1072 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 1073 |
+
if indices is None and cu_seqlens is None and max_seqlen is None:
|
| 1074 |
+
if batch_size is None and seq_len is None:
|
| 1075 |
+
if inputs_embeds is not None:
|
| 1076 |
+
batch_size, seq_len = inputs_embeds.shape[:2]
|
| 1077 |
+
else:
|
| 1078 |
+
batch_size, seq_len = input_ids.shape[:2]
|
| 1079 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
| 1080 |
+
|
| 1081 |
+
if attention_mask is None:
|
| 1082 |
+
attention_mask = torch.ones((batch_size, seq_len), device=device, dtype=torch.bool)
|
| 1083 |
+
|
| 1084 |
+
if inputs_embeds is None:
|
| 1085 |
+
with torch.no_grad():
|
| 1086 |
+
input_ids, indices, cu_seqlens, max_seqlen, position_ids, labels = _unpad_modernbert_input(
|
| 1087 |
+
inputs=input_ids, attention_mask=attention_mask, position_ids=position_ids, labels=labels
|
| 1088 |
+
)
|
| 1089 |
+
else:
|
| 1090 |
+
inputs_embeds, indices, cu_seqlens, max_seqlen, position_ids, labels = _unpad_modernbert_input(
|
| 1091 |
+
inputs=inputs_embeds, attention_mask=attention_mask, position_ids=position_ids, labels=labels
|
| 1092 |
+
)
|
| 1093 |
+
|
| 1094 |
+
outputs = self.model(
|
| 1095 |
+
input_ids=input_ids,
|
| 1096 |
+
attention_mask=attention_mask,
|
| 1097 |
+
sliding_window_mask=sliding_window_mask,
|
| 1098 |
+
position_ids=position_ids,
|
| 1099 |
+
inputs_embeds=inputs_embeds,
|
| 1100 |
+
indices=indices,
|
| 1101 |
+
cu_seqlens=cu_seqlens,
|
| 1102 |
+
max_seqlen=max_seqlen,
|
| 1103 |
+
batch_size=batch_size,
|
| 1104 |
+
seq_len=seq_len,
|
| 1105 |
+
output_attentions=output_attentions,
|
| 1106 |
+
output_hidden_states=output_hidden_states,
|
| 1107 |
+
return_dict=return_dict,
|
| 1108 |
+
)
|
| 1109 |
+
last_hidden_state = outputs[0]
|
| 1110 |
+
|
| 1111 |
+
if self.sparse_prediction and labels is not None:
|
| 1112 |
+
# flatten labels and output first
|
| 1113 |
+
labels = labels.view(-1)
|
| 1114 |
+
last_hidden_state = last_hidden_state.view(labels.shape[0], -1)
|
| 1115 |
+
|
| 1116 |
+
# then filter out the non-masked tokens
|
| 1117 |
+
mask_tokens = labels != self.sparse_pred_ignore_index
|
| 1118 |
+
last_hidden_state = last_hidden_state[mask_tokens]
|
| 1119 |
+
labels = labels[mask_tokens]
|
| 1120 |
+
|
| 1121 |
+
logits = (
|
| 1122 |
+
self.compiled_head(last_hidden_state)
|
| 1123 |
+
if self.config.reference_compile
|
| 1124 |
+
else self.decoder(self.head(last_hidden_state))
|
| 1125 |
+
)
|
| 1126 |
+
|
| 1127 |
+
loss = None
|
| 1128 |
+
if labels is not None:
|
| 1129 |
+
loss = self.loss_function(logits, labels, vocab_size=self.config.vocab_size)
|
| 1130 |
+
|
| 1131 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 1132 |
+
with torch.no_grad():
|
| 1133 |
+
logits = _pad_modernbert_output(inputs=logits, indices=indices, batch=batch_size, seqlen=seq_len)
|
| 1134 |
+
if not return_dict:
|
| 1135 |
+
output = (logits,)
|
| 1136 |
+
return ((loss,) + output) if loss is not None else output
|
| 1137 |
+
|
| 1138 |
+
return MaskedLMOutput(
|
| 1139 |
+
loss=loss,
|
| 1140 |
+
logits=logits,
|
| 1141 |
+
hidden_states=outputs.hidden_states,
|
| 1142 |
+
attentions=outputs.attentions,
|
| 1143 |
+
)
|
| 1144 |
+
|
| 1145 |
+
|
| 1146 |
+
@add_start_docstrings(
|
| 1147 |
+
"The ModernBert Model with a sequence classification head on top that performs pooling.",
|
| 1148 |
+
MODERNBERT_START_DOCSTRING,
|
| 1149 |
+
)
|
| 1150 |
+
class ModernBertForSequenceClassification(ModernBertPreTrainedModel):
|
| 1151 |
+
def __init__(self, config: ModernBertConfig):
|
| 1152 |
+
super().__init__(config)
|
| 1153 |
+
self.num_labels = config.num_labels
|
| 1154 |
+
self.config = config
|
| 1155 |
+
|
| 1156 |
+
self.model = ModernBertModel(config)
|
| 1157 |
+
self.head = ModernBertPredictionHead(config)
|
| 1158 |
+
self.drop = torch.nn.Dropout(config.classifier_dropout)
|
| 1159 |
+
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
| 1160 |
+
|
| 1161 |
+
# Initialize weights and apply final processing
|
| 1162 |
+
self.post_init()
|
| 1163 |
+
|
| 1164 |
+
@add_start_docstrings_to_model_forward(MODERNBERT_INPUTS_DOCSTRING)
|
| 1165 |
+
@add_code_sample_docstrings(
|
| 1166 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1167 |
+
output_type=SequenceClassifierOutput,
|
| 1168 |
+
config_class=_CONFIG_FOR_DOC,
|
| 1169 |
+
)
|
| 1170 |
+
def forward(
|
| 1171 |
+
self,
|
| 1172 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1173 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1174 |
+
sliding_window_mask: Optional[torch.Tensor] = None,
|
| 1175 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1176 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1177 |
+
labels: Optional[torch.Tensor] = None,
|
| 1178 |
+
indices: Optional[torch.Tensor] = None,
|
| 1179 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 1180 |
+
max_seqlen: Optional[int] = None,
|
| 1181 |
+
batch_size: Optional[int] = None,
|
| 1182 |
+
seq_len: Optional[int] = None,
|
| 1183 |
+
output_attentions: Optional[bool] = None,
|
| 1184 |
+
output_hidden_states: Optional[bool] = None,
|
| 1185 |
+
return_dict: Optional[bool] = None,
|
| 1186 |
+
**kwargs,
|
| 1187 |
+
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
|
| 1188 |
+
r"""
|
| 1189 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1190 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1191 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1192 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1193 |
+
"""
|
| 1194 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1195 |
+
self._maybe_set_compile()
|
| 1196 |
+
|
| 1197 |
+
outputs = self.model(
|
| 1198 |
+
input_ids=input_ids,
|
| 1199 |
+
attention_mask=attention_mask,
|
| 1200 |
+
sliding_window_mask=sliding_window_mask,
|
| 1201 |
+
position_ids=position_ids,
|
| 1202 |
+
inputs_embeds=inputs_embeds,
|
| 1203 |
+
indices=indices,
|
| 1204 |
+
cu_seqlens=cu_seqlens,
|
| 1205 |
+
max_seqlen=max_seqlen,
|
| 1206 |
+
batch_size=batch_size,
|
| 1207 |
+
seq_len=seq_len,
|
| 1208 |
+
output_attentions=output_attentions,
|
| 1209 |
+
output_hidden_states=output_hidden_states,
|
| 1210 |
+
return_dict=return_dict,
|
| 1211 |
+
)
|
| 1212 |
+
last_hidden_state = outputs[0]
|
| 1213 |
+
|
| 1214 |
+
if self.config.classifier_pooling == "cls":
|
| 1215 |
+
last_hidden_state = last_hidden_state[:, 0]
|
| 1216 |
+
elif self.config.classifier_pooling == "mean":
|
| 1217 |
+
last_hidden_state = (last_hidden_state * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(
|
| 1218 |
+
dim=1, keepdim=True
|
| 1219 |
+
)
|
| 1220 |
+
|
| 1221 |
+
pooled_output = self.head(last_hidden_state)
|
| 1222 |
+
pooled_output = self.drop(pooled_output)
|
| 1223 |
+
logits = self.classifier(pooled_output)
|
| 1224 |
+
|
| 1225 |
+
loss = None
|
| 1226 |
+
if labels is not None:
|
| 1227 |
+
if self.config.problem_type is None:
|
| 1228 |
+
if self.num_labels == 1:
|
| 1229 |
+
self.config.problem_type = "regression"
|
| 1230 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 1231 |
+
self.config.problem_type = "single_label_classification"
|
| 1232 |
+
else:
|
| 1233 |
+
self.config.problem_type = "multi_label_classification"
|
| 1234 |
+
|
| 1235 |
+
if self.config.problem_type == "regression":
|
| 1236 |
+
loss_fct = MSELoss()
|
| 1237 |
+
if self.num_labels == 1:
|
| 1238 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 1239 |
+
else:
|
| 1240 |
+
loss = loss_fct(logits, labels)
|
| 1241 |
+
elif self.config.problem_type == "single_label_classification":
|
| 1242 |
+
loss_fct = CrossEntropyLoss()
|
| 1243 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 1244 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 1245 |
+
loss_fct = BCEWithLogitsLoss()
|
| 1246 |
+
loss = loss_fct(logits, labels)
|
| 1247 |
+
|
| 1248 |
+
if not return_dict:
|
| 1249 |
+
output = (logits,)
|
| 1250 |
+
return ((loss,) + output) if loss is not None else output
|
| 1251 |
+
|
| 1252 |
+
return SequenceClassifierOutput(
|
| 1253 |
+
loss=loss,
|
| 1254 |
+
logits=logits,
|
| 1255 |
+
hidden_states=outputs.hidden_states,
|
| 1256 |
+
attentions=outputs.attentions,
|
| 1257 |
+
)
|
| 1258 |
+
|
| 1259 |
+
|
| 1260 |
+
@add_start_docstrings(
|
| 1261 |
+
"The ModernBert Model with a token classification head on top, e.g. for Named Entity Recognition (NER) tasks.",
|
| 1262 |
+
MODERNBERT_START_DOCSTRING,
|
| 1263 |
+
)
|
| 1264 |
+
class ModernBertForTokenClassification(ModernBertPreTrainedModel):
|
| 1265 |
+
def __init__(self, config: ModernBertConfig):
|
| 1266 |
+
super().__init__(config)
|
| 1267 |
+
self.num_labels = config.num_labels
|
| 1268 |
+
|
| 1269 |
+
self.model = ModernBertModel(config)
|
| 1270 |
+
self.head = ModernBertPredictionHead(config)
|
| 1271 |
+
self.drop = torch.nn.Dropout(config.classifier_dropout)
|
| 1272 |
+
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
| 1273 |
+
|
| 1274 |
+
# Initialize weights and apply final processing
|
| 1275 |
+
self.post_init()
|
| 1276 |
+
|
| 1277 |
+
@add_start_docstrings_to_model_forward(MODERNBERT_INPUTS_DOCSTRING)
|
| 1278 |
+
@add_code_sample_docstrings(
|
| 1279 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1280 |
+
output_type=TokenClassifierOutput,
|
| 1281 |
+
config_class=_CONFIG_FOR_DOC,
|
| 1282 |
+
)
|
| 1283 |
+
def forward(
|
| 1284 |
+
self,
|
| 1285 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1286 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1287 |
+
sliding_window_mask: Optional[torch.Tensor] = None,
|
| 1288 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1289 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1290 |
+
labels: Optional[torch.Tensor] = None,
|
| 1291 |
+
indices: Optional[torch.Tensor] = None,
|
| 1292 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 1293 |
+
max_seqlen: Optional[int] = None,
|
| 1294 |
+
batch_size: Optional[int] = None,
|
| 1295 |
+
seq_len: Optional[int] = None,
|
| 1296 |
+
output_attentions: Optional[bool] = None,
|
| 1297 |
+
output_hidden_states: Optional[bool] = None,
|
| 1298 |
+
return_dict: Optional[bool] = None,
|
| 1299 |
+
) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
|
| 1300 |
+
r"""
|
| 1301 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1302 |
+
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
|
| 1303 |
+
"""
|
| 1304 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1305 |
+
self._maybe_set_compile()
|
| 1306 |
+
|
| 1307 |
+
outputs = self.model(
|
| 1308 |
+
input_ids=input_ids,
|
| 1309 |
+
attention_mask=attention_mask,
|
| 1310 |
+
sliding_window_mask=sliding_window_mask,
|
| 1311 |
+
position_ids=position_ids,
|
| 1312 |
+
inputs_embeds=inputs_embeds,
|
| 1313 |
+
indices=indices,
|
| 1314 |
+
cu_seqlens=cu_seqlens,
|
| 1315 |
+
max_seqlen=max_seqlen,
|
| 1316 |
+
batch_size=batch_size,
|
| 1317 |
+
seq_len=seq_len,
|
| 1318 |
+
output_attentions=output_attentions,
|
| 1319 |
+
output_hidden_states=output_hidden_states,
|
| 1320 |
+
return_dict=return_dict,
|
| 1321 |
+
)
|
| 1322 |
+
last_hidden_state = outputs[0]
|
| 1323 |
+
|
| 1324 |
+
last_hidden_state = self.head(last_hidden_state)
|
| 1325 |
+
last_hidden_state = self.drop(last_hidden_state)
|
| 1326 |
+
logits = self.classifier(last_hidden_state)
|
| 1327 |
+
|
| 1328 |
+
loss = None
|
| 1329 |
+
if labels is not None:
|
| 1330 |
+
loss_fct = CrossEntropyLoss()
|
| 1331 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 1332 |
+
|
| 1333 |
+
if not return_dict:
|
| 1334 |
+
output = (logits,) + outputs[1:]
|
| 1335 |
+
return ((loss,) + output) if loss is not None else output
|
| 1336 |
+
|
| 1337 |
+
return TokenClassifierOutput(
|
| 1338 |
+
loss=loss,
|
| 1339 |
+
logits=logits,
|
| 1340 |
+
hidden_states=outputs.hidden_states,
|
| 1341 |
+
attentions=outputs.attentions,
|
| 1342 |
+
)
|
| 1343 |
+
|
| 1344 |
+
|
| 1345 |
+
__all__ = [
|
| 1346 |
+
"ModernBertModel",
|
| 1347 |
+
"ModernBertPreTrainedModel",
|
| 1348 |
+
"ModernBertForMaskedLM",
|
| 1349 |
+
"ModernBertForSequenceClassification",
|
| 1350 |
+
"ModernBertForTokenClassification",
|
| 1351 |
+
]
|
pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:efe2082c2825cec0a29fe7b584ef7f631f597e82cf6d3efae7fe46fa900041aa
|
| 3 |
+
size 121641506
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": {
|
| 3 |
+
"content": "[CLS]",
|
| 4 |
+
"lstrip": false,
|
| 5 |
+
"normalized": false,
|
| 6 |
+
"rstrip": false,
|
| 7 |
+
"single_word": false
|
| 8 |
+
},
|
| 9 |
+
"cls_token": {
|
| 10 |
+
"content": "[CLS]",
|
| 11 |
+
"lstrip": false,
|
| 12 |
+
"normalized": false,
|
| 13 |
+
"rstrip": false,
|
| 14 |
+
"single_word": false
|
| 15 |
+
},
|
| 16 |
+
"eos_token": {
|
| 17 |
+
"content": "[SEP]",
|
| 18 |
+
"lstrip": false,
|
| 19 |
+
"normalized": false,
|
| 20 |
+
"rstrip": false,
|
| 21 |
+
"single_word": false
|
| 22 |
+
},
|
| 23 |
+
"mask_token": {
|
| 24 |
+
"content": "[MASK]",
|
| 25 |
+
"lstrip": false,
|
| 26 |
+
"normalized": false,
|
| 27 |
+
"rstrip": false,
|
| 28 |
+
"single_word": false
|
| 29 |
+
},
|
| 30 |
+
"pad_token": {
|
| 31 |
+
"content": "[PAD]",
|
| 32 |
+
"lstrip": false,
|
| 33 |
+
"normalized": false,
|
| 34 |
+
"rstrip": false,
|
| 35 |
+
"single_word": false
|
| 36 |
+
},
|
| 37 |
+
"sep_token": {
|
| 38 |
+
"content": "[SEP]",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false
|
| 43 |
+
},
|
| 44 |
+
"unk_token": {
|
| 45 |
+
"content": "[UNK]",
|
| 46 |
+
"lstrip": false,
|
| 47 |
+
"normalized": false,
|
| 48 |
+
"rstrip": false,
|
| 49 |
+
"single_word": false
|
| 50 |
+
}
|
| 51 |
+
}
|
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": "[CLS]",
|
| 5 |
+
"lstrip": false,
|
| 6 |
+
"normalized": false,
|
| 7 |
+
"rstrip": false,
|
| 8 |
+
"single_word": false,
|
| 9 |
+
"special": true
|
| 10 |
+
},
|
| 11 |
+
"1": {
|
| 12 |
+
"content": "[PAD]",
|
| 13 |
+
"lstrip": false,
|
| 14 |
+
"normalized": false,
|
| 15 |
+
"rstrip": false,
|
| 16 |
+
"single_word": false,
|
| 17 |
+
"special": true
|
| 18 |
+
},
|
| 19 |
+
"2": {
|
| 20 |
+
"content": "[SEP]",
|
| 21 |
+
"lstrip": false,
|
| 22 |
+
"normalized": false,
|
| 23 |
+
"rstrip": false,
|
| 24 |
+
"single_word": false,
|
| 25 |
+
"special": true
|
| 26 |
+
},
|
| 27 |
+
"3": {
|
| 28 |
+
"content": "[UNK]",
|
| 29 |
+
"lstrip": false,
|
| 30 |
+
"normalized": false,
|
| 31 |
+
"rstrip": false,
|
| 32 |
+
"single_word": false,
|
| 33 |
+
"special": true
|
| 34 |
+
},
|
| 35 |
+
"4": {
|
| 36 |
+
"content": "[MASK]",
|
| 37 |
+
"lstrip": false,
|
| 38 |
+
"normalized": false,
|
| 39 |
+
"rstrip": false,
|
| 40 |
+
"single_word": false,
|
| 41 |
+
"special": true
|
| 42 |
+
}
|
| 43 |
+
},
|
| 44 |
+
"bos_token": "[CLS]",
|
| 45 |
+
"clean_up_tokenization_spaces": false,
|
| 46 |
+
"cls_token": "[CLS]",
|
| 47 |
+
"do_lower_case": false,
|
| 48 |
+
"eos_token": "[SEP]",
|
| 49 |
+
"extra_special_tokens": {},
|
| 50 |
+
"keep_accents": true,
|
| 51 |
+
"mask_token": "[MASK]",
|
| 52 |
+
"model_input_names": [
|
| 53 |
+
"input_ids",
|
| 54 |
+
"attention_mask"
|
| 55 |
+
],
|
| 56 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 57 |
+
"pad_token": "[PAD]",
|
| 58 |
+
"sep_token": "[SEP]",
|
| 59 |
+
"split_by_punct": true,
|
| 60 |
+
"tokenizer_class": "DebertaV2TokenizerFast",
|
| 61 |
+
"unk_token": "[UNK]"
|
| 62 |
+
}
|
training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bd3079607bf4955436100ad0ac0e404efb15b7865f1228bdab677a4123d0d337
|
| 3 |
+
size 5368
|
ud.py
ADDED
|
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy
|
| 2 |
+
from transformers import TokenClassificationPipeline
|
| 3 |
+
|
| 4 |
+
class UniversalDependenciesPipeline(TokenClassificationPipeline):
|
| 5 |
+
def __init__(self,**kwargs):
|
| 6 |
+
super().__init__(**kwargs)
|
| 7 |
+
x=self.model.config.label2id
|
| 8 |
+
self.root=numpy.full((len(x)),-numpy.inf)
|
| 9 |
+
self.arc=numpy.full((len(x)),-numpy.inf)
|
| 10 |
+
for k,v in x.items():
|
| 11 |
+
if k.endswith("|[root]"):
|
| 12 |
+
self.root[v]=0
|
| 13 |
+
elif k.endswith("]"):
|
| 14 |
+
self.arc[v]=0
|
| 15 |
+
def _forward(self,model_inputs):
|
| 16 |
+
import torch
|
| 17 |
+
v=model_inputs["input_ids"][0].tolist()
|
| 18 |
+
if len(v)<91:
|
| 19 |
+
x=[True]*(len(v)-2)
|
| 20 |
+
else:
|
| 21 |
+
with torch.no_grad():
|
| 22 |
+
e=self.model(input_ids=torch.tensor([v]).to(self.device))
|
| 23 |
+
m=e.logits[0].cpu().numpy()
|
| 24 |
+
e=numpy.exp(m-numpy.max(m,axis=-1,keepdims=True))
|
| 25 |
+
z=e/e.sum(axis=-1,keepdims=True)
|
| 26 |
+
k=numpy.argmax(m,axis=1).tolist()
|
| 27 |
+
x=[not self.model.config.id2label[p].split("|")[0].endswith(".") for p in k[1:-1]]
|
| 28 |
+
w=(sum([1 for b in x if b])+1)*(len(x)+1)+1
|
| 29 |
+
for i in numpy.argsort([z[i+1,k[i+1]] for i in range(len(x))]):
|
| 30 |
+
if w+len(x)>8191:
|
| 31 |
+
break
|
| 32 |
+
if not x[i]:
|
| 33 |
+
x[i]=True
|
| 34 |
+
w+=len(x)+1
|
| 35 |
+
ids=list(v)
|
| 36 |
+
for i in range(len(x)):
|
| 37 |
+
if x[i]:
|
| 38 |
+
ids+=v[1:i+1]+[self.tokenizer.mask_token_id]+v[i+2:]
|
| 39 |
+
with torch.no_grad():
|
| 40 |
+
e=self.model(input_ids=torch.tensor([ids]).to(self.device))
|
| 41 |
+
return {"logits":e.logits,"thin_out":x,**model_inputs}
|
| 42 |
+
def check_model_type(self,supported_models):
|
| 43 |
+
pass
|
| 44 |
+
def postprocess(self,model_outputs,**kwargs):
|
| 45 |
+
if "logits" not in model_outputs:
|
| 46 |
+
return "".join(self.postprocess(x,**kwargs) for x in model_outputs)
|
| 47 |
+
m=model_outputs["logits"][0].cpu().numpy()
|
| 48 |
+
x=model_outputs["thin_out"]
|
| 49 |
+
e=numpy.full((len(x),len(x),m.shape[-1]),m.min())
|
| 50 |
+
k=len(x)+2
|
| 51 |
+
for i in range(len(x)):
|
| 52 |
+
if x[i]:
|
| 53 |
+
for j in range(len(x)):
|
| 54 |
+
if i==j:
|
| 55 |
+
e[i,i]=m[k]+self.root
|
| 56 |
+
else:
|
| 57 |
+
e[i,j]=m[k]+self.arc
|
| 58 |
+
k+=1
|
| 59 |
+
k+=1
|
| 60 |
+
g=self.model.config.label2id["X.|[goeswith]"]
|
| 61 |
+
m,r=numpy.max(e,axis=2),numpy.tri(e.shape[0])
|
| 62 |
+
for i in range(e.shape[0]):
|
| 63 |
+
for j in range(i+2,e.shape[1]):
|
| 64 |
+
r[i,j]=1
|
| 65 |
+
if numpy.argmax(e[i,j-1])==g:
|
| 66 |
+
if numpy.argmax(m[:,j-1])==i:
|
| 67 |
+
r[i,j]=r[i,j-1]
|
| 68 |
+
e[:,:,g]+=numpy.where(r==0,0,-numpy.inf)
|
| 69 |
+
m,p=numpy.max(e,axis=2),numpy.argmax(e,axis=2)
|
| 70 |
+
h=self.chu_liu_edmonds(m)
|
| 71 |
+
z=[i for i,j in enumerate(h) if i==j]
|
| 72 |
+
if len(z)>1:
|
| 73 |
+
k,h=z[numpy.argmax(m[z,z])],numpy.min(m)-numpy.max(m)
|
| 74 |
+
m[:,z]+=[[0 if j in z and (i!=j or i==k) else h for i in z] for j in range(m.shape[0])]
|
| 75 |
+
h=self.chu_liu_edmonds(m)
|
| 76 |
+
v=[(s,e) for s,e in model_outputs["offset_mapping"][0].tolist() if s<e]
|
| 77 |
+
q=[self.model.config.id2label[p[j,i]].split("|") for i,j in enumerate(h)]
|
| 78 |
+
if "aggregation_strategy" in kwargs and kwargs["aggregation_strategy"]!="none":
|
| 79 |
+
for i,j in reversed(list(enumerate(q[1:],1))):
|
| 80 |
+
if j[-1]=="[goeswith]" and set([t[-1] for t in q[h[i]+1:i+1]])=={"[goeswith]"}:
|
| 81 |
+
h=[b if i>b else b-1 for a,b in enumerate(h) if i!=a]
|
| 82 |
+
v[i-1]=(v[i-1][0],v.pop(i)[1])
|
| 83 |
+
q.pop(i)
|
| 84 |
+
elif v[i-1][1]>v[i][0]:
|
| 85 |
+
h=[b if i>b else b-1 for a,b in enumerate(h) if i!=a]
|
| 86 |
+
v[i-1]=(v[i-1][0],v.pop(i)[1])
|
| 87 |
+
q.pop(i)
|
| 88 |
+
t=model_outputs["sentence"].replace("\n"," ")
|
| 89 |
+
u="# text = "+t+"\n"
|
| 90 |
+
for i,(s,e) in enumerate(v):
|
| 91 |
+
u+="\t".join([str(i+1),t[s:e],"_",q[i][0].replace(".",""),"_","_" if len(q[i])<3 else "|".join(q[i][1:-1]),str(0 if h[i]==i else h[i]+1),q[i][-1][1:-1],"_","_" if i+1<len(v) and e<v[i+1][0] else "SpaceAfter=No"])+"\n"
|
| 92 |
+
return u+"\n"
|
| 93 |
+
def chu_liu_edmonds(self,matrix):
|
| 94 |
+
import numpy
|
| 95 |
+
h=numpy.argmax(matrix,axis=0)
|
| 96 |
+
x=[-1 if i==j else j for i,j in enumerate(h)]
|
| 97 |
+
for b in [lambda x,i,j:-1 if i not in x else x[i],lambda x,i,j:-1 if j<0 else x[j]]:
|
| 98 |
+
y=[]
|
| 99 |
+
while x!=y:
|
| 100 |
+
y=list(x)
|
| 101 |
+
for i,j in enumerate(x):
|
| 102 |
+
x[i]=b(x,i,j)
|
| 103 |
+
if max(x)<0:
|
| 104 |
+
return h
|
| 105 |
+
y,x=[i for i,j in enumerate(x) if j==max(x)],[i for i,j in enumerate(x) if j<max(x)]
|
| 106 |
+
z=matrix-numpy.max(matrix,axis=0)
|
| 107 |
+
m=numpy.block([[z[x,:][:,x],numpy.max(z[x,:][:,y],axis=1).reshape(len(x),1)],[numpy.max(z[y,:][:,x],axis=0),numpy.max(z[y,y])]])
|
| 108 |
+
k=[j if i==len(x) else x[j] if j<len(x) else y[numpy.argmax(z[y,x[i]])] for i,j in enumerate(self.chu_liu_edmonds(m))]
|
| 109 |
+
h=[j if i in y else k[x.index(i)] for i,j in enumerate(h)]
|
| 110 |
+
i=y[numpy.argmax(z[x[k[-1]],y] if k[-1]<len(x) else z[y,y])]
|
| 111 |
+
h[i]=x[k[-1]] if k[-1]<len(x) else i
|
| 112 |
+
return h
|