alex-ht commited on
Commit
9acaf43
1 Parent(s): 1b64a30

first commit

Browse files
README.md ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - fixie-ai/librispeech_asr
4
+ - fixie-ai/common_voice_17_0
5
+ - fixie-ai/peoples_speech
6
+ - fixie-ai/gigaspeech
7
+ - fixie-ai/multilingual_librispeech
8
+ - fixie-ai/wenetspeech
9
+ - fixie-ai/covost2
10
+ language:
11
+ - ar
12
+ - de
13
+ - en
14
+ - es
15
+ - fr
16
+ - hi
17
+ - it
18
+ - ja
19
+ - nl
20
+ - pt
21
+ - ru
22
+ - sv
23
+ - tr
24
+ - uk
25
+ - zh
26
+ library_name: transformers
27
+ license: mit
28
+ metrics:
29
+ - bleu
30
+ ---
31
+
32
+ # Model Card for Ultravox
33
+
34
+ Ultravox is a multimodal Speech LLM built around a pretrained [Llama3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B) and [whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo) backbone.
35
+
36
+ See https://ultravox.ai for the GitHub repo and more information.
37
+
38
+
39
+ ## Model Details
40
+
41
+ ### Model Description
42
+
43
+ Ultravox is a multimodal model that can consume both speech and text as input (e.g., a text system prompt and voice user message).
44
+ The input to the model is given as a text prompt with a special `<|audio|>` pseudo-token, and the model processor will replace this magic token with embeddings derived from the input audio.
45
+ Using the merged embeddings as input, the model will then generate output text as usual.
46
+
47
+ In a future revision of Ultravox, we plan to expand the token vocabulary to support generation of semantic and acoustic audio tokens, which can then be fed to a vocoder to produce voice output.
48
+ No preference tuning has been applied to this revision of the model.
49
+
50
+ - **Developed by:** Fixie.ai
51
+ - **License:** MIT
52
+
53
+ ### Model Sources
54
+
55
+ - **Repository:** https://ultravox.ai
56
+ - **Demo:** See repo
57
+
58
+ ## Usage
59
+
60
+ Think of the model as an LLM that can also hear and understand speech. As such, it can be used as a voice agent, and also to do speech-to-speech translation, analysis of spoken audio, etc.
61
+
62
+ To use the model, try the following:
63
+ ```python
64
+ # pip install transformers peft librosa
65
+
66
+ import transformers
67
+ import numpy as np
68
+ import librosa
69
+
70
+ pipe = transformers.pipeline(model='fixie-ai/ultravox-v0_4_1-llama-3_1-8b', trust_remote_code=True)
71
+
72
+ path = "<path-to-input-audio>" # TODO: pass the audio here
73
+ audio, sr = librosa.load(path, sr=16000)
74
+
75
+
76
+ turns = [
77
+ {
78
+ "role": "system",
79
+ "content": "You are a friendly and helpful character. You love to answer questions for people."
80
+ },
81
+ ]
82
+ pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=30)
83
+ ```
84
+
85
+
86
+ ## Training Details
87
+
88
+ The model uses a pre-trained [Llama3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B) backbone as well as the encoder part of [whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo).
89
+
90
+ Only the multi-modal adapter is trained, while Whisper encoder and Llama are kept frozen.
91
+
92
+ We use a knowledge-distillation loss where Ultravox is trying to match the logits of the text-based Llama backbone.
93
+
94
+ ### Training Data
95
+
96
+ The training dataset is a mix of ASR datasets, extended with continuations generated by Llama 3.1 8B, and speech translation datasets, which yield a modest improvement in translation evaluations.
97
+
98
+ ### Training Procedure
99
+
100
+ Supervised speech instruction finetuning via knowledge-distillation. For more info, see [training code in Ultravox repo](https://github.com/fixie-ai/ultravox/blob/main/ultravox/training/train.py).
101
+
102
+
103
+ #### Training Hyperparameters
104
+
105
+ - **Training regime:** BF16 mixed precision training
106
+ - **Hardward used:** 8x H100 GPUs
107
+
108
+ #### Speeds, Sizes, Times
109
+
110
+ The current version of Ultravox, when invoked with audio content, has a time-to-first-token (TTFT) of approximately 150ms, and a tokens-per-second rate of ~50-100 when using an A100-40GB GPU, all using a Llama 3.1 8B backbone.
111
+
112
+ Check out the audio tab on [TheFastest.ai](https://thefastest.ai/?m=audio) for daily benchmarks and a comparison with other existing models.
113
+
114
+ ## Evaluation
115
+
116
+ | | Ultravox 0.4 8B | **Ultravox 0.4.1 8B** |
117
+ | --- | ---: | ---: |
118
+ | **en_ar** | 11.17 | 12.28 |
119
+ | **en_de** | 25.47 | 27.13 |
120
+ | **es_en** | 37.11 | 39.16 |
121
+ | **ru_en** | 38.96 | 39.65 |
122
+ | **en_ca** | 27.46 | 29.94 |
123
+ | **zh_en** | 10.08 | 14.55 |
config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "AlexHung29629/test_mllama_v12",
3
+ "architectures": [
4
+ "UltravoxModel"
5
+ ],
6
+ "audio_model_id": "openai/whisper-large-v3-turbo",
7
+ "auto_map": {
8
+ "AutoConfig": "ultravox_config.UltravoxConfig",
9
+ "AutoModel": "ultravox_model.UltravoxModel",
10
+ "AutoProcessor": "ultravox_processing.UltravoxProcessor"
11
+ },
12
+ "custom_pipelines": {
13
+ "ultravox-pipeline": {
14
+ "impl": "ultravox_pipeline.UltravoxPipeline",
15
+ "pt": [
16
+ "AutoModel"
17
+ ],
18
+ "tf": [],
19
+ "type": "multimodal"
20
+ }
21
+ },
22
+ "hidden_size": 4096,
23
+ "ignore_index": -100,
24
+ "initializer_range": 0.02,
25
+ "model_type": "ultravox",
26
+ "norm_init": 0.4,
27
+ "pad_token_id": 128009,
28
+ "projector_act": "swiglu",
29
+ "stack_factor": 8,
30
+ "text_model_id": "voidful/Llama-3.2-8B-Instruct",
31
+ "torch_dtype": "bfloat16",
32
+ "transformers_version": "4.46.2",
33
+ "vocab_size": 128256
34
+ }
generation_config.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 128000,
4
+ "eos_token_id": [
5
+ 128001,
6
+ 128008,
7
+ 128009
8
+ ],
9
+ "pad_token_id": 128009,
10
+ "transformers_version": "4.46.2"
11
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb49be3de02b6f00f4e698145d8ab74c6b487cfc26686ddc161d4353b908c99d
3
+ size 100692448
special_tokens_map.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<|begin_of_text|>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<|eot_id|>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<|eot_id|>"
17
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,2063 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "added_tokens_decoder": {
3
+ "128000": {
4
+ "content": "<|begin_of_text|>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false,
9
+ "special": true
10
+ },
11
+ "128001": {
12
+ "content": "<|end_of_text|>",
13
+ "lstrip": false,
14
+ "normalized": false,
15
+ "rstrip": false,
16
+ "single_word": false,
17
+ "special": true
18
+ },
19
+ "128002": {
20
+ "content": "<|reserved_special_token_0|>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false,
25
+ "special": true
26
+ },
27
+ "128003": {
28
+ "content": "<|reserved_special_token_1|>",
29
+ "lstrip": false,
30
+ "normalized": false,
31
+ "rstrip": false,
32
+ "single_word": false,
33
+ "special": true
34
+ },
35
+ "128004": {
36
+ "content": "<|finetune_right_pad_id|>",
37
+ "lstrip": false,
38
+ "normalized": false,
39
+ "rstrip": false,
40
+ "single_word": false,
41
+ "special": true
42
+ },
43
+ "128005": {
44
+ "content": "<|reserved_special_token_2|>",
45
+ "lstrip": false,
46
+ "normalized": false,
47
+ "rstrip": false,
48
+ "single_word": false,
49
+ "special": true
50
+ },
51
+ "128006": {
52
+ "content": "<|start_header_id|>",
53
+ "lstrip": false,
54
+ "normalized": false,
55
+ "rstrip": false,
56
+ "single_word": false,
57
+ "special": true
58
+ },
59
+ "128007": {
60
+ "content": "<|end_header_id|>",
61
+ "lstrip": false,
62
+ "normalized": false,
63
+ "rstrip": false,
64
+ "single_word": false,
65
+ "special": true
66
+ },
67
+ "128008": {
68
+ "content": "<|eom_id|>",
69
+ "lstrip": false,
70
+ "normalized": false,
71
+ "rstrip": false,
72
+ "single_word": false,
73
+ "special": true
74
+ },
75
+ "128009": {
76
+ "content": "<|eot_id|>",
77
+ "lstrip": false,
78
+ "normalized": false,
79
+ "rstrip": false,
80
+ "single_word": false,
81
+ "special": true
82
+ },
83
+ "128010": {
84
+ "content": "<|python_tag|>",
85
+ "lstrip": false,
86
+ "normalized": false,
87
+ "rstrip": false,
88
+ "single_word": false,
89
+ "special": true
90
+ },
91
+ "128011": {
92
+ "content": "<|reserved_special_token_3|>",
93
+ "lstrip": false,
94
+ "normalized": false,
95
+ "rstrip": false,
96
+ "single_word": false,
97
+ "special": true
98
+ },
99
+ "128012": {
100
+ "content": "<|reserved_special_token_4|>",
101
+ "lstrip": false,
102
+ "normalized": false,
103
+ "rstrip": false,
104
+ "single_word": false,
105
+ "special": true
106
+ },
107
+ "128013": {
108
+ "content": "<|reserved_special_token_5|>",
109
+ "lstrip": false,
110
+ "normalized": false,
111
+ "rstrip": false,
112
+ "single_word": false,
113
+ "special": true
114
+ },
115
+ "128014": {
116
+ "content": "<|reserved_special_token_6|>",
117
+ "lstrip": false,
118
+ "normalized": false,
119
+ "rstrip": false,
120
+ "single_word": false,
121
+ "special": true
122
+ },
123
+ "128015": {
124
+ "content": "<|reserved_special_token_7|>",
125
+ "lstrip": false,
126
+ "normalized": false,
127
+ "rstrip": false,
128
+ "single_word": false,
129
+ "special": true
130
+ },
131
+ "128016": {
132
+ "content": "<|reserved_special_token_8|>",
133
+ "lstrip": false,
134
+ "normalized": false,
135
+ "rstrip": false,
136
+ "single_word": false,
137
+ "special": true
138
+ },
139
+ "128017": {
140
+ "content": "<|reserved_special_token_9|>",
141
+ "lstrip": false,
142
+ "normalized": false,
143
+ "rstrip": false,
144
+ "single_word": false,
145
+ "special": true
146
+ },
147
+ "128018": {
148
+ "content": "<|reserved_special_token_10|>",
149
+ "lstrip": false,
150
+ "normalized": false,
151
+ "rstrip": false,
152
+ "single_word": false,
153
+ "special": true
154
+ },
155
+ "128019": {
156
+ "content": "<|reserved_special_token_11|>",
157
+ "lstrip": false,
158
+ "normalized": false,
159
+ "rstrip": false,
160
+ "single_word": false,
161
+ "special": true
162
+ },
163
+ "128020": {
164
+ "content": "<|reserved_special_token_12|>",
165
+ "lstrip": false,
166
+ "normalized": false,
167
+ "rstrip": false,
168
+ "single_word": false,
169
+ "special": true
170
+ },
171
+ "128021": {
172
+ "content": "<|reserved_special_token_13|>",
173
+ "lstrip": false,
174
+ "normalized": false,
175
+ "rstrip": false,
176
+ "single_word": false,
177
+ "special": true
178
+ },
179
+ "128022": {
180
+ "content": "<|reserved_special_token_14|>",
181
+ "lstrip": false,
182
+ "normalized": false,
183
+ "rstrip": false,
184
+ "single_word": false,
185
+ "special": true
186
+ },
187
+ "128023": {
188
+ "content": "<|reserved_special_token_15|>",
189
+ "lstrip": false,
190
+ "normalized": false,
191
+ "rstrip": false,
192
+ "single_word": false,
193
+ "special": true
194
+ },
195
+ "128024": {
196
+ "content": "<|reserved_special_token_16|>",
197
+ "lstrip": false,
198
+ "normalized": false,
199
+ "rstrip": false,
200
+ "single_word": false,
201
+ "special": true
202
+ },
203
+ "128025": {
204
+ "content": "<|reserved_special_token_17|>",
205
+ "lstrip": false,
206
+ "normalized": false,
207
+ "rstrip": false,
208
+ "single_word": false,
209
+ "special": true
210
+ },
211
+ "128026": {
212
+ "content": "<|reserved_special_token_18|>",
213
+ "lstrip": false,
214
+ "normalized": false,
215
+ "rstrip": false,
216
+ "single_word": false,
217
+ "special": true
218
+ },
219
+ "128027": {
220
+ "content": "<|reserved_special_token_19|>",
221
+ "lstrip": false,
222
+ "normalized": false,
223
+ "rstrip": false,
224
+ "single_word": false,
225
+ "special": true
226
+ },
227
+ "128028": {
228
+ "content": "<|reserved_special_token_20|>",
229
+ "lstrip": false,
230
+ "normalized": false,
231
+ "rstrip": false,
232
+ "single_word": false,
233
+ "special": true
234
+ },
235
+ "128029": {
236
+ "content": "<|reserved_special_token_21|>",
237
+ "lstrip": false,
238
+ "normalized": false,
239
+ "rstrip": false,
240
+ "single_word": false,
241
+ "special": true
242
+ },
243
+ "128030": {
244
+ "content": "<|reserved_special_token_22|>",
245
+ "lstrip": false,
246
+ "normalized": false,
247
+ "rstrip": false,
248
+ "single_word": false,
249
+ "special": true
250
+ },
251
+ "128031": {
252
+ "content": "<|reserved_special_token_23|>",
253
+ "lstrip": false,
254
+ "normalized": false,
255
+ "rstrip": false,
256
+ "single_word": false,
257
+ "special": true
258
+ },
259
+ "128032": {
260
+ "content": "<|reserved_special_token_24|>",
261
+ "lstrip": false,
262
+ "normalized": false,
263
+ "rstrip": false,
264
+ "single_word": false,
265
+ "special": true
266
+ },
267
+ "128033": {
268
+ "content": "<|reserved_special_token_25|>",
269
+ "lstrip": false,
270
+ "normalized": false,
271
+ "rstrip": false,
272
+ "single_word": false,
273
+ "special": true
274
+ },
275
+ "128034": {
276
+ "content": "<|reserved_special_token_26|>",
277
+ "lstrip": false,
278
+ "normalized": false,
279
+ "rstrip": false,
280
+ "single_word": false,
281
+ "special": true
282
+ },
283
+ "128035": {
284
+ "content": "<|reserved_special_token_27|>",
285
+ "lstrip": false,
286
+ "normalized": false,
287
+ "rstrip": false,
288
+ "single_word": false,
289
+ "special": true
290
+ },
291
+ "128036": {
292
+ "content": "<|reserved_special_token_28|>",
293
+ "lstrip": false,
294
+ "normalized": false,
295
+ "rstrip": false,
296
+ "single_word": false,
297
+ "special": true
298
+ },
299
+ "128037": {
300
+ "content": "<|reserved_special_token_29|>",
301
+ "lstrip": false,
302
+ "normalized": false,
303
+ "rstrip": false,
304
+ "single_word": false,
305
+ "special": true
306
+ },
307
+ "128038": {
308
+ "content": "<|reserved_special_token_30|>",
309
+ "lstrip": false,
310
+ "normalized": false,
311
+ "rstrip": false,
312
+ "single_word": false,
313
+ "special": true
314
+ },
315
+ "128039": {
316
+ "content": "<|reserved_special_token_31|>",
317
+ "lstrip": false,
318
+ "normalized": false,
319
+ "rstrip": false,
320
+ "single_word": false,
321
+ "special": true
322
+ },
323
+ "128040": {
324
+ "content": "<|reserved_special_token_32|>",
325
+ "lstrip": false,
326
+ "normalized": false,
327
+ "rstrip": false,
328
+ "single_word": false,
329
+ "special": true
330
+ },
331
+ "128041": {
332
+ "content": "<|reserved_special_token_33|>",
333
+ "lstrip": false,
334
+ "normalized": false,
335
+ "rstrip": false,
336
+ "single_word": false,
337
+ "special": true
338
+ },
339
+ "128042": {
340
+ "content": "<|reserved_special_token_34|>",
341
+ "lstrip": false,
342
+ "normalized": false,
343
+ "rstrip": false,
344
+ "single_word": false,
345
+ "special": true
346
+ },
347
+ "128043": {
348
+ "content": "<|reserved_special_token_35|>",
349
+ "lstrip": false,
350
+ "normalized": false,
351
+ "rstrip": false,
352
+ "single_word": false,
353
+ "special": true
354
+ },
355
+ "128044": {
356
+ "content": "<|reserved_special_token_36|>",
357
+ "lstrip": false,
358
+ "normalized": false,
359
+ "rstrip": false,
360
+ "single_word": false,
361
+ "special": true
362
+ },
363
+ "128045": {
364
+ "content": "<|reserved_special_token_37|>",
365
+ "lstrip": false,
366
+ "normalized": false,
367
+ "rstrip": false,
368
+ "single_word": false,
369
+ "special": true
370
+ },
371
+ "128046": {
372
+ "content": "<|reserved_special_token_38|>",
373
+ "lstrip": false,
374
+ "normalized": false,
375
+ "rstrip": false,
376
+ "single_word": false,
377
+ "special": true
378
+ },
379
+ "128047": {
380
+ "content": "<|reserved_special_token_39|>",
381
+ "lstrip": false,
382
+ "normalized": false,
383
+ "rstrip": false,
384
+ "single_word": false,
385
+ "special": true
386
+ },
387
+ "128048": {
388
+ "content": "<|reserved_special_token_40|>",
389
+ "lstrip": false,
390
+ "normalized": false,
391
+ "rstrip": false,
392
+ "single_word": false,
393
+ "special": true
394
+ },
395
+ "128049": {
396
+ "content": "<|reserved_special_token_41|>",
397
+ "lstrip": false,
398
+ "normalized": false,
399
+ "rstrip": false,
400
+ "single_word": false,
401
+ "special": true
402
+ },
403
+ "128050": {
404
+ "content": "<|reserved_special_token_42|>",
405
+ "lstrip": false,
406
+ "normalized": false,
407
+ "rstrip": false,
408
+ "single_word": false,
409
+ "special": true
410
+ },
411
+ "128051": {
412
+ "content": "<|reserved_special_token_43|>",
413
+ "lstrip": false,
414
+ "normalized": false,
415
+ "rstrip": false,
416
+ "single_word": false,
417
+ "special": true
418
+ },
419
+ "128052": {
420
+ "content": "<|reserved_special_token_44|>",
421
+ "lstrip": false,
422
+ "normalized": false,
423
+ "rstrip": false,
424
+ "single_word": false,
425
+ "special": true
426
+ },
427
+ "128053": {
428
+ "content": "<|reserved_special_token_45|>",
429
+ "lstrip": false,
430
+ "normalized": false,
431
+ "rstrip": false,
432
+ "single_word": false,
433
+ "special": true
434
+ },
435
+ "128054": {
436
+ "content": "<|reserved_special_token_46|>",
437
+ "lstrip": false,
438
+ "normalized": false,
439
+ "rstrip": false,
440
+ "single_word": false,
441
+ "special": true
442
+ },
443
+ "128055": {
444
+ "content": "<|reserved_special_token_47|>",
445
+ "lstrip": false,
446
+ "normalized": false,
447
+ "rstrip": false,
448
+ "single_word": false,
449
+ "special": true
450
+ },
451
+ "128056": {
452
+ "content": "<|reserved_special_token_48|>",
453
+ "lstrip": false,
454
+ "normalized": false,
455
+ "rstrip": false,
456
+ "single_word": false,
457
+ "special": true
458
+ },
459
+ "128057": {
460
+ "content": "<|reserved_special_token_49|>",
461
+ "lstrip": false,
462
+ "normalized": false,
463
+ "rstrip": false,
464
+ "single_word": false,
465
+ "special": true
466
+ },
467
+ "128058": {
468
+ "content": "<|reserved_special_token_50|>",
469
+ "lstrip": false,
470
+ "normalized": false,
471
+ "rstrip": false,
472
+ "single_word": false,
473
+ "special": true
474
+ },
475
+ "128059": {
476
+ "content": "<|reserved_special_token_51|>",
477
+ "lstrip": false,
478
+ "normalized": false,
479
+ "rstrip": false,
480
+ "single_word": false,
481
+ "special": true
482
+ },
483
+ "128060": {
484
+ "content": "<|reserved_special_token_52|>",
485
+ "lstrip": false,
486
+ "normalized": false,
487
+ "rstrip": false,
488
+ "single_word": false,
489
+ "special": true
490
+ },
491
+ "128061": {
492
+ "content": "<|reserved_special_token_53|>",
493
+ "lstrip": false,
494
+ "normalized": false,
495
+ "rstrip": false,
496
+ "single_word": false,
497
+ "special": true
498
+ },
499
+ "128062": {
500
+ "content": "<|reserved_special_token_54|>",
501
+ "lstrip": false,
502
+ "normalized": false,
503
+ "rstrip": false,
504
+ "single_word": false,
505
+ "special": true
506
+ },
507
+ "128063": {
508
+ "content": "<|reserved_special_token_55|>",
509
+ "lstrip": false,
510
+ "normalized": false,
511
+ "rstrip": false,
512
+ "single_word": false,
513
+ "special": true
514
+ },
515
+ "128064": {
516
+ "content": "<|reserved_special_token_56|>",
517
+ "lstrip": false,
518
+ "normalized": false,
519
+ "rstrip": false,
520
+ "single_word": false,
521
+ "special": true
522
+ },
523
+ "128065": {
524
+ "content": "<|reserved_special_token_57|>",
525
+ "lstrip": false,
526
+ "normalized": false,
527
+ "rstrip": false,
528
+ "single_word": false,
529
+ "special": true
530
+ },
531
+ "128066": {
532
+ "content": "<|reserved_special_token_58|>",
533
+ "lstrip": false,
534
+ "normalized": false,
535
+ "rstrip": false,
536
+ "single_word": false,
537
+ "special": true
538
+ },
539
+ "128067": {
540
+ "content": "<|reserved_special_token_59|>",
541
+ "lstrip": false,
542
+ "normalized": false,
543
+ "rstrip": false,
544
+ "single_word": false,
545
+ "special": true
546
+ },
547
+ "128068": {
548
+ "content": "<|reserved_special_token_60|>",
549
+ "lstrip": false,
550
+ "normalized": false,
551
+ "rstrip": false,
552
+ "single_word": false,
553
+ "special": true
554
+ },
555
+ "128069": {
556
+ "content": "<|reserved_special_token_61|>",
557
+ "lstrip": false,
558
+ "normalized": false,
559
+ "rstrip": false,
560
+ "single_word": false,
561
+ "special": true
562
+ },
563
+ "128070": {
564
+ "content": "<|reserved_special_token_62|>",
565
+ "lstrip": false,
566
+ "normalized": false,
567
+ "rstrip": false,
568
+ "single_word": false,
569
+ "special": true
570
+ },
571
+ "128071": {
572
+ "content": "<|reserved_special_token_63|>",
573
+ "lstrip": false,
574
+ "normalized": false,
575
+ "rstrip": false,
576
+ "single_word": false,
577
+ "special": true
578
+ },
579
+ "128072": {
580
+ "content": "<|reserved_special_token_64|>",
581
+ "lstrip": false,
582
+ "normalized": false,
583
+ "rstrip": false,
584
+ "single_word": false,
585
+ "special": true
586
+ },
587
+ "128073": {
588
+ "content": "<|reserved_special_token_65|>",
589
+ "lstrip": false,
590
+ "normalized": false,
591
+ "rstrip": false,
592
+ "single_word": false,
593
+ "special": true
594
+ },
595
+ "128074": {
596
+ "content": "<|reserved_special_token_66|>",
597
+ "lstrip": false,
598
+ "normalized": false,
599
+ "rstrip": false,
600
+ "single_word": false,
601
+ "special": true
602
+ },
603
+ "128075": {
604
+ "content": "<|reserved_special_token_67|>",
605
+ "lstrip": false,
606
+ "normalized": false,
607
+ "rstrip": false,
608
+ "single_word": false,
609
+ "special": true
610
+ },
611
+ "128076": {
612
+ "content": "<|reserved_special_token_68|>",
613
+ "lstrip": false,
614
+ "normalized": false,
615
+ "rstrip": false,
616
+ "single_word": false,
617
+ "special": true
618
+ },
619
+ "128077": {
620
+ "content": "<|reserved_special_token_69|>",
621
+ "lstrip": false,
622
+ "normalized": false,
623
+ "rstrip": false,
624
+ "single_word": false,
625
+ "special": true
626
+ },
627
+ "128078": {
628
+ "content": "<|reserved_special_token_70|>",
629
+ "lstrip": false,
630
+ "normalized": false,
631
+ "rstrip": false,
632
+ "single_word": false,
633
+ "special": true
634
+ },
635
+ "128079": {
636
+ "content": "<|reserved_special_token_71|>",
637
+ "lstrip": false,
638
+ "normalized": false,
639
+ "rstrip": false,
640
+ "single_word": false,
641
+ "special": true
642
+ },
643
+ "128080": {
644
+ "content": "<|reserved_special_token_72|>",
645
+ "lstrip": false,
646
+ "normalized": false,
647
+ "rstrip": false,
648
+ "single_word": false,
649
+ "special": true
650
+ },
651
+ "128081": {
652
+ "content": "<|reserved_special_token_73|>",
653
+ "lstrip": false,
654
+ "normalized": false,
655
+ "rstrip": false,
656
+ "single_word": false,
657
+ "special": true
658
+ },
659
+ "128082": {
660
+ "content": "<|reserved_special_token_74|>",
661
+ "lstrip": false,
662
+ "normalized": false,
663
+ "rstrip": false,
664
+ "single_word": false,
665
+ "special": true
666
+ },
667
+ "128083": {
668
+ "content": "<|reserved_special_token_75|>",
669
+ "lstrip": false,
670
+ "normalized": false,
671
+ "rstrip": false,
672
+ "single_word": false,
673
+ "special": true
674
+ },
675
+ "128084": {
676
+ "content": "<|reserved_special_token_76|>",
677
+ "lstrip": false,
678
+ "normalized": false,
679
+ "rstrip": false,
680
+ "single_word": false,
681
+ "special": true
682
+ },
683
+ "128085": {
684
+ "content": "<|reserved_special_token_77|>",
685
+ "lstrip": false,
686
+ "normalized": false,
687
+ "rstrip": false,
688
+ "single_word": false,
689
+ "special": true
690
+ },
691
+ "128086": {
692
+ "content": "<|reserved_special_token_78|>",
693
+ "lstrip": false,
694
+ "normalized": false,
695
+ "rstrip": false,
696
+ "single_word": false,
697
+ "special": true
698
+ },
699
+ "128087": {
700
+ "content": "<|reserved_special_token_79|>",
701
+ "lstrip": false,
702
+ "normalized": false,
703
+ "rstrip": false,
704
+ "single_word": false,
705
+ "special": true
706
+ },
707
+ "128088": {
708
+ "content": "<|reserved_special_token_80|>",
709
+ "lstrip": false,
710
+ "normalized": false,
711
+ "rstrip": false,
712
+ "single_word": false,
713
+ "special": true
714
+ },
715
+ "128089": {
716
+ "content": "<|reserved_special_token_81|>",
717
+ "lstrip": false,
718
+ "normalized": false,
719
+ "rstrip": false,
720
+ "single_word": false,
721
+ "special": true
722
+ },
723
+ "128090": {
724
+ "content": "<|reserved_special_token_82|>",
725
+ "lstrip": false,
726
+ "normalized": false,
727
+ "rstrip": false,
728
+ "single_word": false,
729
+ "special": true
730
+ },
731
+ "128091": {
732
+ "content": "<|reserved_special_token_83|>",
733
+ "lstrip": false,
734
+ "normalized": false,
735
+ "rstrip": false,
736
+ "single_word": false,
737
+ "special": true
738
+ },
739
+ "128092": {
740
+ "content": "<|reserved_special_token_84|>",
741
+ "lstrip": false,
742
+ "normalized": false,
743
+ "rstrip": false,
744
+ "single_word": false,
745
+ "special": true
746
+ },
747
+ "128093": {
748
+ "content": "<|reserved_special_token_85|>",
749
+ "lstrip": false,
750
+ "normalized": false,
751
+ "rstrip": false,
752
+ "single_word": false,
753
+ "special": true
754
+ },
755
+ "128094": {
756
+ "content": "<|reserved_special_token_86|>",
757
+ "lstrip": false,
758
+ "normalized": false,
759
+ "rstrip": false,
760
+ "single_word": false,
761
+ "special": true
762
+ },
763
+ "128095": {
764
+ "content": "<|reserved_special_token_87|>",
765
+ "lstrip": false,
766
+ "normalized": false,
767
+ "rstrip": false,
768
+ "single_word": false,
769
+ "special": true
770
+ },
771
+ "128096": {
772
+ "content": "<|reserved_special_token_88|>",
773
+ "lstrip": false,
774
+ "normalized": false,
775
+ "rstrip": false,
776
+ "single_word": false,
777
+ "special": true
778
+ },
779
+ "128097": {
780
+ "content": "<|reserved_special_token_89|>",
781
+ "lstrip": false,
782
+ "normalized": false,
783
+ "rstrip": false,
784
+ "single_word": false,
785
+ "special": true
786
+ },
787
+ "128098": {
788
+ "content": "<|reserved_special_token_90|>",
789
+ "lstrip": false,
790
+ "normalized": false,
791
+ "rstrip": false,
792
+ "single_word": false,
793
+ "special": true
794
+ },
795
+ "128099": {
796
+ "content": "<|reserved_special_token_91|>",
797
+ "lstrip": false,
798
+ "normalized": false,
799
+ "rstrip": false,
800
+ "single_word": false,
801
+ "special": true
802
+ },
803
+ "128100": {
804
+ "content": "<|reserved_special_token_92|>",
805
+ "lstrip": false,
806
+ "normalized": false,
807
+ "rstrip": false,
808
+ "single_word": false,
809
+ "special": true
810
+ },
811
+ "128101": {
812
+ "content": "<|reserved_special_token_93|>",
813
+ "lstrip": false,
814
+ "normalized": false,
815
+ "rstrip": false,
816
+ "single_word": false,
817
+ "special": true
818
+ },
819
+ "128102": {
820
+ "content": "<|reserved_special_token_94|>",
821
+ "lstrip": false,
822
+ "normalized": false,
823
+ "rstrip": false,
824
+ "single_word": false,
825
+ "special": true
826
+ },
827
+ "128103": {
828
+ "content": "<|reserved_special_token_95|>",
829
+ "lstrip": false,
830
+ "normalized": false,
831
+ "rstrip": false,
832
+ "single_word": false,
833
+ "special": true
834
+ },
835
+ "128104": {
836
+ "content": "<|reserved_special_token_96|>",
837
+ "lstrip": false,
838
+ "normalized": false,
839
+ "rstrip": false,
840
+ "single_word": false,
841
+ "special": true
842
+ },
843
+ "128105": {
844
+ "content": "<|reserved_special_token_97|>",
845
+ "lstrip": false,
846
+ "normalized": false,
847
+ "rstrip": false,
848
+ "single_word": false,
849
+ "special": true
850
+ },
851
+ "128106": {
852
+ "content": "<|reserved_special_token_98|>",
853
+ "lstrip": false,
854
+ "normalized": false,
855
+ "rstrip": false,
856
+ "single_word": false,
857
+ "special": true
858
+ },
859
+ "128107": {
860
+ "content": "<|reserved_special_token_99|>",
861
+ "lstrip": false,
862
+ "normalized": false,
863
+ "rstrip": false,
864
+ "single_word": false,
865
+ "special": true
866
+ },
867
+ "128108": {
868
+ "content": "<|reserved_special_token_100|>",
869
+ "lstrip": false,
870
+ "normalized": false,
871
+ "rstrip": false,
872
+ "single_word": false,
873
+ "special": true
874
+ },
875
+ "128109": {
876
+ "content": "<|reserved_special_token_101|>",
877
+ "lstrip": false,
878
+ "normalized": false,
879
+ "rstrip": false,
880
+ "single_word": false,
881
+ "special": true
882
+ },
883
+ "128110": {
884
+ "content": "<|reserved_special_token_102|>",
885
+ "lstrip": false,
886
+ "normalized": false,
887
+ "rstrip": false,
888
+ "single_word": false,
889
+ "special": true
890
+ },
891
+ "128111": {
892
+ "content": "<|reserved_special_token_103|>",
893
+ "lstrip": false,
894
+ "normalized": false,
895
+ "rstrip": false,
896
+ "single_word": false,
897
+ "special": true
898
+ },
899
+ "128112": {
900
+ "content": "<|reserved_special_token_104|>",
901
+ "lstrip": false,
902
+ "normalized": false,
903
+ "rstrip": false,
904
+ "single_word": false,
905
+ "special": true
906
+ },
907
+ "128113": {
908
+ "content": "<|reserved_special_token_105|>",
909
+ "lstrip": false,
910
+ "normalized": false,
911
+ "rstrip": false,
912
+ "single_word": false,
913
+ "special": true
914
+ },
915
+ "128114": {
916
+ "content": "<|reserved_special_token_106|>",
917
+ "lstrip": false,
918
+ "normalized": false,
919
+ "rstrip": false,
920
+ "single_word": false,
921
+ "special": true
922
+ },
923
+ "128115": {
924
+ "content": "<|reserved_special_token_107|>",
925
+ "lstrip": false,
926
+ "normalized": false,
927
+ "rstrip": false,
928
+ "single_word": false,
929
+ "special": true
930
+ },
931
+ "128116": {
932
+ "content": "<|reserved_special_token_108|>",
933
+ "lstrip": false,
934
+ "normalized": false,
935
+ "rstrip": false,
936
+ "single_word": false,
937
+ "special": true
938
+ },
939
+ "128117": {
940
+ "content": "<|reserved_special_token_109|>",
941
+ "lstrip": false,
942
+ "normalized": false,
943
+ "rstrip": false,
944
+ "single_word": false,
945
+ "special": true
946
+ },
947
+ "128118": {
948
+ "content": "<|reserved_special_token_110|>",
949
+ "lstrip": false,
950
+ "normalized": false,
951
+ "rstrip": false,
952
+ "single_word": false,
953
+ "special": true
954
+ },
955
+ "128119": {
956
+ "content": "<|reserved_special_token_111|>",
957
+ "lstrip": false,
958
+ "normalized": false,
959
+ "rstrip": false,
960
+ "single_word": false,
961
+ "special": true
962
+ },
963
+ "128120": {
964
+ "content": "<|reserved_special_token_112|>",
965
+ "lstrip": false,
966
+ "normalized": false,
967
+ "rstrip": false,
968
+ "single_word": false,
969
+ "special": true
970
+ },
971
+ "128121": {
972
+ "content": "<|reserved_special_token_113|>",
973
+ "lstrip": false,
974
+ "normalized": false,
975
+ "rstrip": false,
976
+ "single_word": false,
977
+ "special": true
978
+ },
979
+ "128122": {
980
+ "content": "<|reserved_special_token_114|>",
981
+ "lstrip": false,
982
+ "normalized": false,
983
+ "rstrip": false,
984
+ "single_word": false,
985
+ "special": true
986
+ },
987
+ "128123": {
988
+ "content": "<|reserved_special_token_115|>",
989
+ "lstrip": false,
990
+ "normalized": false,
991
+ "rstrip": false,
992
+ "single_word": false,
993
+ "special": true
994
+ },
995
+ "128124": {
996
+ "content": "<|reserved_special_token_116|>",
997
+ "lstrip": false,
998
+ "normalized": false,
999
+ "rstrip": false,
1000
+ "single_word": false,
1001
+ "special": true
1002
+ },
1003
+ "128125": {
1004
+ "content": "<|reserved_special_token_117|>",
1005
+ "lstrip": false,
1006
+ "normalized": false,
1007
+ "rstrip": false,
1008
+ "single_word": false,
1009
+ "special": true
1010
+ },
1011
+ "128126": {
1012
+ "content": "<|reserved_special_token_118|>",
1013
+ "lstrip": false,
1014
+ "normalized": false,
1015
+ "rstrip": false,
1016
+ "single_word": false,
1017
+ "special": true
1018
+ },
1019
+ "128127": {
1020
+ "content": "<|reserved_special_token_119|>",
1021
+ "lstrip": false,
1022
+ "normalized": false,
1023
+ "rstrip": false,
1024
+ "single_word": false,
1025
+ "special": true
1026
+ },
1027
+ "128128": {
1028
+ "content": "<|reserved_special_token_120|>",
1029
+ "lstrip": false,
1030
+ "normalized": false,
1031
+ "rstrip": false,
1032
+ "single_word": false,
1033
+ "special": true
1034
+ },
1035
+ "128129": {
1036
+ "content": "<|reserved_special_token_121|>",
1037
+ "lstrip": false,
1038
+ "normalized": false,
1039
+ "rstrip": false,
1040
+ "single_word": false,
1041
+ "special": true
1042
+ },
1043
+ "128130": {
1044
+ "content": "<|reserved_special_token_122|>",
1045
+ "lstrip": false,
1046
+ "normalized": false,
1047
+ "rstrip": false,
1048
+ "single_word": false,
1049
+ "special": true
1050
+ },
1051
+ "128131": {
1052
+ "content": "<|reserved_special_token_123|>",
1053
+ "lstrip": false,
1054
+ "normalized": false,
1055
+ "rstrip": false,
1056
+ "single_word": false,
1057
+ "special": true
1058
+ },
1059
+ "128132": {
1060
+ "content": "<|reserved_special_token_124|>",
1061
+ "lstrip": false,
1062
+ "normalized": false,
1063
+ "rstrip": false,
1064
+ "single_word": false,
1065
+ "special": true
1066
+ },
1067
+ "128133": {
1068
+ "content": "<|reserved_special_token_125|>",
1069
+ "lstrip": false,
1070
+ "normalized": false,
1071
+ "rstrip": false,
1072
+ "single_word": false,
1073
+ "special": true
1074
+ },
1075
+ "128134": {
1076
+ "content": "<|reserved_special_token_126|>",
1077
+ "lstrip": false,
1078
+ "normalized": false,
1079
+ "rstrip": false,
1080
+ "single_word": false,
1081
+ "special": true
1082
+ },
1083
+ "128135": {
1084
+ "content": "<|reserved_special_token_127|>",
1085
+ "lstrip": false,
1086
+ "normalized": false,
1087
+ "rstrip": false,
1088
+ "single_word": false,
1089
+ "special": true
1090
+ },
1091
+ "128136": {
1092
+ "content": "<|reserved_special_token_128|>",
1093
+ "lstrip": false,
1094
+ "normalized": false,
1095
+ "rstrip": false,
1096
+ "single_word": false,
1097
+ "special": true
1098
+ },
1099
+ "128137": {
1100
+ "content": "<|reserved_special_token_129|>",
1101
+ "lstrip": false,
1102
+ "normalized": false,
1103
+ "rstrip": false,
1104
+ "single_word": false,
1105
+ "special": true
1106
+ },
1107
+ "128138": {
1108
+ "content": "<|reserved_special_token_130|>",
1109
+ "lstrip": false,
1110
+ "normalized": false,
1111
+ "rstrip": false,
1112
+ "single_word": false,
1113
+ "special": true
1114
+ },
1115
+ "128139": {
1116
+ "content": "<|reserved_special_token_131|>",
1117
+ "lstrip": false,
1118
+ "normalized": false,
1119
+ "rstrip": false,
1120
+ "single_word": false,
1121
+ "special": true
1122
+ },
1123
+ "128140": {
1124
+ "content": "<|reserved_special_token_132|>",
1125
+ "lstrip": false,
1126
+ "normalized": false,
1127
+ "rstrip": false,
1128
+ "single_word": false,
1129
+ "special": true
1130
+ },
1131
+ "128141": {
1132
+ "content": "<|reserved_special_token_133|>",
1133
+ "lstrip": false,
1134
+ "normalized": false,
1135
+ "rstrip": false,
1136
+ "single_word": false,
1137
+ "special": true
1138
+ },
1139
+ "128142": {
1140
+ "content": "<|reserved_special_token_134|>",
1141
+ "lstrip": false,
1142
+ "normalized": false,
1143
+ "rstrip": false,
1144
+ "single_word": false,
1145
+ "special": true
1146
+ },
1147
+ "128143": {
1148
+ "content": "<|reserved_special_token_135|>",
1149
+ "lstrip": false,
1150
+ "normalized": false,
1151
+ "rstrip": false,
1152
+ "single_word": false,
1153
+ "special": true
1154
+ },
1155
+ "128144": {
1156
+ "content": "<|reserved_special_token_136|>",
1157
+ "lstrip": false,
1158
+ "normalized": false,
1159
+ "rstrip": false,
1160
+ "single_word": false,
1161
+ "special": true
1162
+ },
1163
+ "128145": {
1164
+ "content": "<|reserved_special_token_137|>",
1165
+ "lstrip": false,
1166
+ "normalized": false,
1167
+ "rstrip": false,
1168
+ "single_word": false,
1169
+ "special": true
1170
+ },
1171
+ "128146": {
1172
+ "content": "<|reserved_special_token_138|>",
1173
+ "lstrip": false,
1174
+ "normalized": false,
1175
+ "rstrip": false,
1176
+ "single_word": false,
1177
+ "special": true
1178
+ },
1179
+ "128147": {
1180
+ "content": "<|reserved_special_token_139|>",
1181
+ "lstrip": false,
1182
+ "normalized": false,
1183
+ "rstrip": false,
1184
+ "single_word": false,
1185
+ "special": true
1186
+ },
1187
+ "128148": {
1188
+ "content": "<|reserved_special_token_140|>",
1189
+ "lstrip": false,
1190
+ "normalized": false,
1191
+ "rstrip": false,
1192
+ "single_word": false,
1193
+ "special": true
1194
+ },
1195
+ "128149": {
1196
+ "content": "<|reserved_special_token_141|>",
1197
+ "lstrip": false,
1198
+ "normalized": false,
1199
+ "rstrip": false,
1200
+ "single_word": false,
1201
+ "special": true
1202
+ },
1203
+ "128150": {
1204
+ "content": "<|reserved_special_token_142|>",
1205
+ "lstrip": false,
1206
+ "normalized": false,
1207
+ "rstrip": false,
1208
+ "single_word": false,
1209
+ "special": true
1210
+ },
1211
+ "128151": {
1212
+ "content": "<|reserved_special_token_143|>",
1213
+ "lstrip": false,
1214
+ "normalized": false,
1215
+ "rstrip": false,
1216
+ "single_word": false,
1217
+ "special": true
1218
+ },
1219
+ "128152": {
1220
+ "content": "<|reserved_special_token_144|>",
1221
+ "lstrip": false,
1222
+ "normalized": false,
1223
+ "rstrip": false,
1224
+ "single_word": false,
1225
+ "special": true
1226
+ },
1227
+ "128153": {
1228
+ "content": "<|reserved_special_token_145|>",
1229
+ "lstrip": false,
1230
+ "normalized": false,
1231
+ "rstrip": false,
1232
+ "single_word": false,
1233
+ "special": true
1234
+ },
1235
+ "128154": {
1236
+ "content": "<|reserved_special_token_146|>",
1237
+ "lstrip": false,
1238
+ "normalized": false,
1239
+ "rstrip": false,
1240
+ "single_word": false,
1241
+ "special": true
1242
+ },
1243
+ "128155": {
1244
+ "content": "<|reserved_special_token_147|>",
1245
+ "lstrip": false,
1246
+ "normalized": false,
1247
+ "rstrip": false,
1248
+ "single_word": false,
1249
+ "special": true
1250
+ },
1251
+ "128156": {
1252
+ "content": "<|reserved_special_token_148|>",
1253
+ "lstrip": false,
1254
+ "normalized": false,
1255
+ "rstrip": false,
1256
+ "single_word": false,
1257
+ "special": true
1258
+ },
1259
+ "128157": {
1260
+ "content": "<|reserved_special_token_149|>",
1261
+ "lstrip": false,
1262
+ "normalized": false,
1263
+ "rstrip": false,
1264
+ "single_word": false,
1265
+ "special": true
1266
+ },
1267
+ "128158": {
1268
+ "content": "<|reserved_special_token_150|>",
1269
+ "lstrip": false,
1270
+ "normalized": false,
1271
+ "rstrip": false,
1272
+ "single_word": false,
1273
+ "special": true
1274
+ },
1275
+ "128159": {
1276
+ "content": "<|reserved_special_token_151|>",
1277
+ "lstrip": false,
1278
+ "normalized": false,
1279
+ "rstrip": false,
1280
+ "single_word": false,
1281
+ "special": true
1282
+ },
1283
+ "128160": {
1284
+ "content": "<|reserved_special_token_152|>",
1285
+ "lstrip": false,
1286
+ "normalized": false,
1287
+ "rstrip": false,
1288
+ "single_word": false,
1289
+ "special": true
1290
+ },
1291
+ "128161": {
1292
+ "content": "<|reserved_special_token_153|>",
1293
+ "lstrip": false,
1294
+ "normalized": false,
1295
+ "rstrip": false,
1296
+ "single_word": false,
1297
+ "special": true
1298
+ },
1299
+ "128162": {
1300
+ "content": "<|reserved_special_token_154|>",
1301
+ "lstrip": false,
1302
+ "normalized": false,
1303
+ "rstrip": false,
1304
+ "single_word": false,
1305
+ "special": true
1306
+ },
1307
+ "128163": {
1308
+ "content": "<|reserved_special_token_155|>",
1309
+ "lstrip": false,
1310
+ "normalized": false,
1311
+ "rstrip": false,
1312
+ "single_word": false,
1313
+ "special": true
1314
+ },
1315
+ "128164": {
1316
+ "content": "<|reserved_special_token_156|>",
1317
+ "lstrip": false,
1318
+ "normalized": false,
1319
+ "rstrip": false,
1320
+ "single_word": false,
1321
+ "special": true
1322
+ },
1323
+ "128165": {
1324
+ "content": "<|reserved_special_token_157|>",
1325
+ "lstrip": false,
1326
+ "normalized": false,
1327
+ "rstrip": false,
1328
+ "single_word": false,
1329
+ "special": true
1330
+ },
1331
+ "128166": {
1332
+ "content": "<|reserved_special_token_158|>",
1333
+ "lstrip": false,
1334
+ "normalized": false,
1335
+ "rstrip": false,
1336
+ "single_word": false,
1337
+ "special": true
1338
+ },
1339
+ "128167": {
1340
+ "content": "<|reserved_special_token_159|>",
1341
+ "lstrip": false,
1342
+ "normalized": false,
1343
+ "rstrip": false,
1344
+ "single_word": false,
1345
+ "special": true
1346
+ },
1347
+ "128168": {
1348
+ "content": "<|reserved_special_token_160|>",
1349
+ "lstrip": false,
1350
+ "normalized": false,
1351
+ "rstrip": false,
1352
+ "single_word": false,
1353
+ "special": true
1354
+ },
1355
+ "128169": {
1356
+ "content": "<|reserved_special_token_161|>",
1357
+ "lstrip": false,
1358
+ "normalized": false,
1359
+ "rstrip": false,
1360
+ "single_word": false,
1361
+ "special": true
1362
+ },
1363
+ "128170": {
1364
+ "content": "<|reserved_special_token_162|>",
1365
+ "lstrip": false,
1366
+ "normalized": false,
1367
+ "rstrip": false,
1368
+ "single_word": false,
1369
+ "special": true
1370
+ },
1371
+ "128171": {
1372
+ "content": "<|reserved_special_token_163|>",
1373
+ "lstrip": false,
1374
+ "normalized": false,
1375
+ "rstrip": false,
1376
+ "single_word": false,
1377
+ "special": true
1378
+ },
1379
+ "128172": {
1380
+ "content": "<|reserved_special_token_164|>",
1381
+ "lstrip": false,
1382
+ "normalized": false,
1383
+ "rstrip": false,
1384
+ "single_word": false,
1385
+ "special": true
1386
+ },
1387
+ "128173": {
1388
+ "content": "<|reserved_special_token_165|>",
1389
+ "lstrip": false,
1390
+ "normalized": false,
1391
+ "rstrip": false,
1392
+ "single_word": false,
1393
+ "special": true
1394
+ },
1395
+ "128174": {
1396
+ "content": "<|reserved_special_token_166|>",
1397
+ "lstrip": false,
1398
+ "normalized": false,
1399
+ "rstrip": false,
1400
+ "single_word": false,
1401
+ "special": true
1402
+ },
1403
+ "128175": {
1404
+ "content": "<|reserved_special_token_167|>",
1405
+ "lstrip": false,
1406
+ "normalized": false,
1407
+ "rstrip": false,
1408
+ "single_word": false,
1409
+ "special": true
1410
+ },
1411
+ "128176": {
1412
+ "content": "<|reserved_special_token_168|>",
1413
+ "lstrip": false,
1414
+ "normalized": false,
1415
+ "rstrip": false,
1416
+ "single_word": false,
1417
+ "special": true
1418
+ },
1419
+ "128177": {
1420
+ "content": "<|reserved_special_token_169|>",
1421
+ "lstrip": false,
1422
+ "normalized": false,
1423
+ "rstrip": false,
1424
+ "single_word": false,
1425
+ "special": true
1426
+ },
1427
+ "128178": {
1428
+ "content": "<|reserved_special_token_170|>",
1429
+ "lstrip": false,
1430
+ "normalized": false,
1431
+ "rstrip": false,
1432
+ "single_word": false,
1433
+ "special": true
1434
+ },
1435
+ "128179": {
1436
+ "content": "<|reserved_special_token_171|>",
1437
+ "lstrip": false,
1438
+ "normalized": false,
1439
+ "rstrip": false,
1440
+ "single_word": false,
1441
+ "special": true
1442
+ },
1443
+ "128180": {
1444
+ "content": "<|reserved_special_token_172|>",
1445
+ "lstrip": false,
1446
+ "normalized": false,
1447
+ "rstrip": false,
1448
+ "single_word": false,
1449
+ "special": true
1450
+ },
1451
+ "128181": {
1452
+ "content": "<|reserved_special_token_173|>",
1453
+ "lstrip": false,
1454
+ "normalized": false,
1455
+ "rstrip": false,
1456
+ "single_word": false,
1457
+ "special": true
1458
+ },
1459
+ "128182": {
1460
+ "content": "<|reserved_special_token_174|>",
1461
+ "lstrip": false,
1462
+ "normalized": false,
1463
+ "rstrip": false,
1464
+ "single_word": false,
1465
+ "special": true
1466
+ },
1467
+ "128183": {
1468
+ "content": "<|reserved_special_token_175|>",
1469
+ "lstrip": false,
1470
+ "normalized": false,
1471
+ "rstrip": false,
1472
+ "single_word": false,
1473
+ "special": true
1474
+ },
1475
+ "128184": {
1476
+ "content": "<|reserved_special_token_176|>",
1477
+ "lstrip": false,
1478
+ "normalized": false,
1479
+ "rstrip": false,
1480
+ "single_word": false,
1481
+ "special": true
1482
+ },
1483
+ "128185": {
1484
+ "content": "<|reserved_special_token_177|>",
1485
+ "lstrip": false,
1486
+ "normalized": false,
1487
+ "rstrip": false,
1488
+ "single_word": false,
1489
+ "special": true
1490
+ },
1491
+ "128186": {
1492
+ "content": "<|reserved_special_token_178|>",
1493
+ "lstrip": false,
1494
+ "normalized": false,
1495
+ "rstrip": false,
1496
+ "single_word": false,
1497
+ "special": true
1498
+ },
1499
+ "128187": {
1500
+ "content": "<|reserved_special_token_179|>",
1501
+ "lstrip": false,
1502
+ "normalized": false,
1503
+ "rstrip": false,
1504
+ "single_word": false,
1505
+ "special": true
1506
+ },
1507
+ "128188": {
1508
+ "content": "<|reserved_special_token_180|>",
1509
+ "lstrip": false,
1510
+ "normalized": false,
1511
+ "rstrip": false,
1512
+ "single_word": false,
1513
+ "special": true
1514
+ },
1515
+ "128189": {
1516
+ "content": "<|reserved_special_token_181|>",
1517
+ "lstrip": false,
1518
+ "normalized": false,
1519
+ "rstrip": false,
1520
+ "single_word": false,
1521
+ "special": true
1522
+ },
1523
+ "128190": {
1524
+ "content": "<|reserved_special_token_182|>",
1525
+ "lstrip": false,
1526
+ "normalized": false,
1527
+ "rstrip": false,
1528
+ "single_word": false,
1529
+ "special": true
1530
+ },
1531
+ "128191": {
1532
+ "content": "<|reserved_special_token_183|>",
1533
+ "lstrip": false,
1534
+ "normalized": false,
1535
+ "rstrip": false,
1536
+ "single_word": false,
1537
+ "special": true
1538
+ },
1539
+ "128192": {
1540
+ "content": "<|reserved_special_token_184|>",
1541
+ "lstrip": false,
1542
+ "normalized": false,
1543
+ "rstrip": false,
1544
+ "single_word": false,
1545
+ "special": true
1546
+ },
1547
+ "128193": {
1548
+ "content": "<|reserved_special_token_185|>",
1549
+ "lstrip": false,
1550
+ "normalized": false,
1551
+ "rstrip": false,
1552
+ "single_word": false,
1553
+ "special": true
1554
+ },
1555
+ "128194": {
1556
+ "content": "<|reserved_special_token_186|>",
1557
+ "lstrip": false,
1558
+ "normalized": false,
1559
+ "rstrip": false,
1560
+ "single_word": false,
1561
+ "special": true
1562
+ },
1563
+ "128195": {
1564
+ "content": "<|reserved_special_token_187|>",
1565
+ "lstrip": false,
1566
+ "normalized": false,
1567
+ "rstrip": false,
1568
+ "single_word": false,
1569
+ "special": true
1570
+ },
1571
+ "128196": {
1572
+ "content": "<|reserved_special_token_188|>",
1573
+ "lstrip": false,
1574
+ "normalized": false,
1575
+ "rstrip": false,
1576
+ "single_word": false,
1577
+ "special": true
1578
+ },
1579
+ "128197": {
1580
+ "content": "<|reserved_special_token_189|>",
1581
+ "lstrip": false,
1582
+ "normalized": false,
1583
+ "rstrip": false,
1584
+ "single_word": false,
1585
+ "special": true
1586
+ },
1587
+ "128198": {
1588
+ "content": "<|reserved_special_token_190|>",
1589
+ "lstrip": false,
1590
+ "normalized": false,
1591
+ "rstrip": false,
1592
+ "single_word": false,
1593
+ "special": true
1594
+ },
1595
+ "128199": {
1596
+ "content": "<|reserved_special_token_191|>",
1597
+ "lstrip": false,
1598
+ "normalized": false,
1599
+ "rstrip": false,
1600
+ "single_word": false,
1601
+ "special": true
1602
+ },
1603
+ "128200": {
1604
+ "content": "<|reserved_special_token_192|>",
1605
+ "lstrip": false,
1606
+ "normalized": false,
1607
+ "rstrip": false,
1608
+ "single_word": false,
1609
+ "special": true
1610
+ },
1611
+ "128201": {
1612
+ "content": "<|reserved_special_token_193|>",
1613
+ "lstrip": false,
1614
+ "normalized": false,
1615
+ "rstrip": false,
1616
+ "single_word": false,
1617
+ "special": true
1618
+ },
1619
+ "128202": {
1620
+ "content": "<|reserved_special_token_194|>",
1621
+ "lstrip": false,
1622
+ "normalized": false,
1623
+ "rstrip": false,
1624
+ "single_word": false,
1625
+ "special": true
1626
+ },
1627
+ "128203": {
1628
+ "content": "<|reserved_special_token_195|>",
1629
+ "lstrip": false,
1630
+ "normalized": false,
1631
+ "rstrip": false,
1632
+ "single_word": false,
1633
+ "special": true
1634
+ },
1635
+ "128204": {
1636
+ "content": "<|reserved_special_token_196|>",
1637
+ "lstrip": false,
1638
+ "normalized": false,
1639
+ "rstrip": false,
1640
+ "single_word": false,
1641
+ "special": true
1642
+ },
1643
+ "128205": {
1644
+ "content": "<|reserved_special_token_197|>",
1645
+ "lstrip": false,
1646
+ "normalized": false,
1647
+ "rstrip": false,
1648
+ "single_word": false,
1649
+ "special": true
1650
+ },
1651
+ "128206": {
1652
+ "content": "<|reserved_special_token_198|>",
1653
+ "lstrip": false,
1654
+ "normalized": false,
1655
+ "rstrip": false,
1656
+ "single_word": false,
1657
+ "special": true
1658
+ },
1659
+ "128207": {
1660
+ "content": "<|reserved_special_token_199|>",
1661
+ "lstrip": false,
1662
+ "normalized": false,
1663
+ "rstrip": false,
1664
+ "single_word": false,
1665
+ "special": true
1666
+ },
1667
+ "128208": {
1668
+ "content": "<|reserved_special_token_200|>",
1669
+ "lstrip": false,
1670
+ "normalized": false,
1671
+ "rstrip": false,
1672
+ "single_word": false,
1673
+ "special": true
1674
+ },
1675
+ "128209": {
1676
+ "content": "<|reserved_special_token_201|>",
1677
+ "lstrip": false,
1678
+ "normalized": false,
1679
+ "rstrip": false,
1680
+ "single_word": false,
1681
+ "special": true
1682
+ },
1683
+ "128210": {
1684
+ "content": "<|reserved_special_token_202|>",
1685
+ "lstrip": false,
1686
+ "normalized": false,
1687
+ "rstrip": false,
1688
+ "single_word": false,
1689
+ "special": true
1690
+ },
1691
+ "128211": {
1692
+ "content": "<|reserved_special_token_203|>",
1693
+ "lstrip": false,
1694
+ "normalized": false,
1695
+ "rstrip": false,
1696
+ "single_word": false,
1697
+ "special": true
1698
+ },
1699
+ "128212": {
1700
+ "content": "<|reserved_special_token_204|>",
1701
+ "lstrip": false,
1702
+ "normalized": false,
1703
+ "rstrip": false,
1704
+ "single_word": false,
1705
+ "special": true
1706
+ },
1707
+ "128213": {
1708
+ "content": "<|reserved_special_token_205|>",
1709
+ "lstrip": false,
1710
+ "normalized": false,
1711
+ "rstrip": false,
1712
+ "single_word": false,
1713
+ "special": true
1714
+ },
1715
+ "128214": {
1716
+ "content": "<|reserved_special_token_206|>",
1717
+ "lstrip": false,
1718
+ "normalized": false,
1719
+ "rstrip": false,
1720
+ "single_word": false,
1721
+ "special": true
1722
+ },
1723
+ "128215": {
1724
+ "content": "<|reserved_special_token_207|>",
1725
+ "lstrip": false,
1726
+ "normalized": false,
1727
+ "rstrip": false,
1728
+ "single_word": false,
1729
+ "special": true
1730
+ },
1731
+ "128216": {
1732
+ "content": "<|reserved_special_token_208|>",
1733
+ "lstrip": false,
1734
+ "normalized": false,
1735
+ "rstrip": false,
1736
+ "single_word": false,
1737
+ "special": true
1738
+ },
1739
+ "128217": {
1740
+ "content": "<|reserved_special_token_209|>",
1741
+ "lstrip": false,
1742
+ "normalized": false,
1743
+ "rstrip": false,
1744
+ "single_word": false,
1745
+ "special": true
1746
+ },
1747
+ "128218": {
1748
+ "content": "<|reserved_special_token_210|>",
1749
+ "lstrip": false,
1750
+ "normalized": false,
1751
+ "rstrip": false,
1752
+ "single_word": false,
1753
+ "special": true
1754
+ },
1755
+ "128219": {
1756
+ "content": "<|reserved_special_token_211|>",
1757
+ "lstrip": false,
1758
+ "normalized": false,
1759
+ "rstrip": false,
1760
+ "single_word": false,
1761
+ "special": true
1762
+ },
1763
+ "128220": {
1764
+ "content": "<|reserved_special_token_212|>",
1765
+ "lstrip": false,
1766
+ "normalized": false,
1767
+ "rstrip": false,
1768
+ "single_word": false,
1769
+ "special": true
1770
+ },
1771
+ "128221": {
1772
+ "content": "<|reserved_special_token_213|>",
1773
+ "lstrip": false,
1774
+ "normalized": false,
1775
+ "rstrip": false,
1776
+ "single_word": false,
1777
+ "special": true
1778
+ },
1779
+ "128222": {
1780
+ "content": "<|reserved_special_token_214|>",
1781
+ "lstrip": false,
1782
+ "normalized": false,
1783
+ "rstrip": false,
1784
+ "single_word": false,
1785
+ "special": true
1786
+ },
1787
+ "128223": {
1788
+ "content": "<|reserved_special_token_215|>",
1789
+ "lstrip": false,
1790
+ "normalized": false,
1791
+ "rstrip": false,
1792
+ "single_word": false,
1793
+ "special": true
1794
+ },
1795
+ "128224": {
1796
+ "content": "<|reserved_special_token_216|>",
1797
+ "lstrip": false,
1798
+ "normalized": false,
1799
+ "rstrip": false,
1800
+ "single_word": false,
1801
+ "special": true
1802
+ },
1803
+ "128225": {
1804
+ "content": "<|reserved_special_token_217|>",
1805
+ "lstrip": false,
1806
+ "normalized": false,
1807
+ "rstrip": false,
1808
+ "single_word": false,
1809
+ "special": true
1810
+ },
1811
+ "128226": {
1812
+ "content": "<|reserved_special_token_218|>",
1813
+ "lstrip": false,
1814
+ "normalized": false,
1815
+ "rstrip": false,
1816
+ "single_word": false,
1817
+ "special": true
1818
+ },
1819
+ "128227": {
1820
+ "content": "<|reserved_special_token_219|>",
1821
+ "lstrip": false,
1822
+ "normalized": false,
1823
+ "rstrip": false,
1824
+ "single_word": false,
1825
+ "special": true
1826
+ },
1827
+ "128228": {
1828
+ "content": "<|reserved_special_token_220|>",
1829
+ "lstrip": false,
1830
+ "normalized": false,
1831
+ "rstrip": false,
1832
+ "single_word": false,
1833
+ "special": true
1834
+ },
1835
+ "128229": {
1836
+ "content": "<|reserved_special_token_221|>",
1837
+ "lstrip": false,
1838
+ "normalized": false,
1839
+ "rstrip": false,
1840
+ "single_word": false,
1841
+ "special": true
1842
+ },
1843
+ "128230": {
1844
+ "content": "<|reserved_special_token_222|>",
1845
+ "lstrip": false,
1846
+ "normalized": false,
1847
+ "rstrip": false,
1848
+ "single_word": false,
1849
+ "special": true
1850
+ },
1851
+ "128231": {
1852
+ "content": "<|reserved_special_token_223|>",
1853
+ "lstrip": false,
1854
+ "normalized": false,
1855
+ "rstrip": false,
1856
+ "single_word": false,
1857
+ "special": true
1858
+ },
1859
+ "128232": {
1860
+ "content": "<|reserved_special_token_224|>",
1861
+ "lstrip": false,
1862
+ "normalized": false,
1863
+ "rstrip": false,
1864
+ "single_word": false,
1865
+ "special": true
1866
+ },
1867
+ "128233": {
1868
+ "content": "<|reserved_special_token_225|>",
1869
+ "lstrip": false,
1870
+ "normalized": false,
1871
+ "rstrip": false,
1872
+ "single_word": false,
1873
+ "special": true
1874
+ },
1875
+ "128234": {
1876
+ "content": "<|reserved_special_token_226|>",
1877
+ "lstrip": false,
1878
+ "normalized": false,
1879
+ "rstrip": false,
1880
+ "single_word": false,
1881
+ "special": true
1882
+ },
1883
+ "128235": {
1884
+ "content": "<|reserved_special_token_227|>",
1885
+ "lstrip": false,
1886
+ "normalized": false,
1887
+ "rstrip": false,
1888
+ "single_word": false,
1889
+ "special": true
1890
+ },
1891
+ "128236": {
1892
+ "content": "<|reserved_special_token_228|>",
1893
+ "lstrip": false,
1894
+ "normalized": false,
1895
+ "rstrip": false,
1896
+ "single_word": false,
1897
+ "special": true
1898
+ },
1899
+ "128237": {
1900
+ "content": "<|reserved_special_token_229|>",
1901
+ "lstrip": false,
1902
+ "normalized": false,
1903
+ "rstrip": false,
1904
+ "single_word": false,
1905
+ "special": true
1906
+ },
1907
+ "128238": {
1908
+ "content": "<|reserved_special_token_230|>",
1909
+ "lstrip": false,
1910
+ "normalized": false,
1911
+ "rstrip": false,
1912
+ "single_word": false,
1913
+ "special": true
1914
+ },
1915
+ "128239": {
1916
+ "content": "<|reserved_special_token_231|>",
1917
+ "lstrip": false,
1918
+ "normalized": false,
1919
+ "rstrip": false,
1920
+ "single_word": false,
1921
+ "special": true
1922
+ },
1923
+ "128240": {
1924
+ "content": "<|reserved_special_token_232|>",
1925
+ "lstrip": false,
1926
+ "normalized": false,
1927
+ "rstrip": false,
1928
+ "single_word": false,
1929
+ "special": true
1930
+ },
1931
+ "128241": {
1932
+ "content": "<|reserved_special_token_233|>",
1933
+ "lstrip": false,
1934
+ "normalized": false,
1935
+ "rstrip": false,
1936
+ "single_word": false,
1937
+ "special": true
1938
+ },
1939
+ "128242": {
1940
+ "content": "<|reserved_special_token_234|>",
1941
+ "lstrip": false,
1942
+ "normalized": false,
1943
+ "rstrip": false,
1944
+ "single_word": false,
1945
+ "special": true
1946
+ },
1947
+ "128243": {
1948
+ "content": "<|reserved_special_token_235|>",
1949
+ "lstrip": false,
1950
+ "normalized": false,
1951
+ "rstrip": false,
1952
+ "single_word": false,
1953
+ "special": true
1954
+ },
1955
+ "128244": {
1956
+ "content": "<|reserved_special_token_236|>",
1957
+ "lstrip": false,
1958
+ "normalized": false,
1959
+ "rstrip": false,
1960
+ "single_word": false,
1961
+ "special": true
1962
+ },
1963
+ "128245": {
1964
+ "content": "<|reserved_special_token_237|>",
1965
+ "lstrip": false,
1966
+ "normalized": false,
1967
+ "rstrip": false,
1968
+ "single_word": false,
1969
+ "special": true
1970
+ },
1971
+ "128246": {
1972
+ "content": "<|reserved_special_token_238|>",
1973
+ "lstrip": false,
1974
+ "normalized": false,
1975
+ "rstrip": false,
1976
+ "single_word": false,
1977
+ "special": true
1978
+ },
1979
+ "128247": {
1980
+ "content": "<|reserved_special_token_239|>",
1981
+ "lstrip": false,
1982
+ "normalized": false,
1983
+ "rstrip": false,
1984
+ "single_word": false,
1985
+ "special": true
1986
+ },
1987
+ "128248": {
1988
+ "content": "<|reserved_special_token_240|>",
1989
+ "lstrip": false,
1990
+ "normalized": false,
1991
+ "rstrip": false,
1992
+ "single_word": false,
1993
+ "special": true
1994
+ },
1995
+ "128249": {
1996
+ "content": "<|reserved_special_token_241|>",
1997
+ "lstrip": false,
1998
+ "normalized": false,
1999
+ "rstrip": false,
2000
+ "single_word": false,
2001
+ "special": true
2002
+ },
2003
+ "128250": {
2004
+ "content": "<|reserved_special_token_242|>",
2005
+ "lstrip": false,
2006
+ "normalized": false,
2007
+ "rstrip": false,
2008
+ "single_word": false,
2009
+ "special": true
2010
+ },
2011
+ "128251": {
2012
+ "content": "<|reserved_special_token_243|>",
2013
+ "lstrip": false,
2014
+ "normalized": false,
2015
+ "rstrip": false,
2016
+ "single_word": false,
2017
+ "special": true
2018
+ },
2019
+ "128252": {
2020
+ "content": "<|reserved_special_token_244|>",
2021
+ "lstrip": false,
2022
+ "normalized": false,
2023
+ "rstrip": false,
2024
+ "single_word": false,
2025
+ "special": true
2026
+ },
2027
+ "128253": {
2028
+ "content": "<|reserved_special_token_245|>",
2029
+ "lstrip": false,
2030
+ "normalized": false,
2031
+ "rstrip": false,
2032
+ "single_word": false,
2033
+ "special": true
2034
+ },
2035
+ "128254": {
2036
+ "content": "<|reserved_special_token_246|>",
2037
+ "lstrip": false,
2038
+ "normalized": false,
2039
+ "rstrip": false,
2040
+ "single_word": false,
2041
+ "special": true
2042
+ },
2043
+ "128255": {
2044
+ "content": "<|reserved_special_token_247|>",
2045
+ "lstrip": false,
2046
+ "normalized": false,
2047
+ "rstrip": false,
2048
+ "single_word": false,
2049
+ "special": true
2050
+ }
2051
+ },
2052
+ "bos_token": "<|begin_of_text|>",
2053
+ "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
2054
+ "clean_up_tokenization_spaces": true,
2055
+ "eos_token": "<|eot_id|>",
2056
+ "model_input_names": [
2057
+ "input_ids",
2058
+ "attention_mask"
2059
+ ],
2060
+ "model_max_length": 131072,
2061
+ "pad_token": "<|finetune_right_pad_id|>",
2062
+ "tokenizer_class": "PreTrainedTokenizerFast"
2063
+ }
ultravox_config.py ADDED
@@ -0,0 +1,166 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import dataclasses
2
+ from enum import Enum
3
+ from typing import Any, Dict, List, Optional
4
+
5
+ import transformers
6
+
7
+
8
+ @dataclasses.dataclass
9
+ class LoraConfigSimplified:
10
+ """
11
+ Low Rank Approximation (LoRA) configuration.
12
+
13
+ Used for language and audio models separately.
14
+ """
15
+
16
+ # The rank of the approximation
17
+ r: int = 0
18
+ lora_alpha: float = 8
19
+ target_modules: Optional[List[str]] = dataclasses.field(
20
+ default_factory=lambda: ["k_proj", "q_proj", "linear_k", "linear_q"]
21
+ )
22
+
23
+
24
+ class LossFunction(str, Enum):
25
+ CrossEntropy = "ce"
26
+ KL_Divergence = "kl"
27
+
28
+
29
+ @dataclasses.dataclass
30
+ class LossConfig:
31
+ loss_function: LossFunction = LossFunction.CrossEntropy
32
+ kl_temperature: float = 2.0
33
+
34
+ @property
35
+ def requires_alt_fields(self):
36
+ return self.loss_function == LossFunction.KL_Divergence
37
+
38
+
39
+ class UltravoxConfig(transformers.PretrainedConfig):
40
+ r"""
41
+ This is the configuration class to store the configuration of a [`UltravoxForConditionalGeneration`]. It is used to instantiate an
42
+ Ultravox model according to the specified arguments, defining the model architecture.
43
+
44
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
45
+ documentation from [`PretrainedConfig`] for more information.
46
+
47
+ Args:
48
+ audio_config (`Wav2Vec2Config`, *optional*):
49
+ Custom audio config or dict
50
+ text_config (`Union[AutoConfig, dict]`, *optional*):
51
+ The config object of the text backbone. Can be any of `LlamaConfig` or `MistralConfig`.
52
+ ignore_index (`int`, *optional*, defaults to -100):
53
+ The ignore index for the loss function.
54
+ audio_token_index (`int`, *optional*, defaults to 32000):
55
+ The audio token index to encode the audio prompt.
56
+ stack_factor (`int`, *optional*, defaults to 8):
57
+ Audio downsampling factor for the multimodal projector.
58
+ norm_init (`float`, *optional*, defaults to 0.4):
59
+ The initialization value for the layer normalization.
60
+ projector_act (`str`, *optional*, defaults to `"swiglu"`):
61
+ The activation function used by the multimodal projector.
62
+ text_model_lora_config (`LoraConfigSimplified`, *optional*):
63
+ The LoRA configuration for finetuning the text model.
64
+ audio_model_lora_config (`LoraConfigSimplified`, *optional*):
65
+ The LoRA configuration for finetuning the audio model.
66
+
67
+
68
+ Example:
69
+
70
+ ```python
71
+ >>> from transformers import UltravoxForConditionalGeneration, Wav2Vec2Config, UltravoxConfig, LlamaConfig
72
+
73
+ >>> # Initializing an audio encoder config
74
+ >>> audio_config = Wav2Vec2Config()
75
+
76
+ >>> # Initializing a Llama config
77
+ >>> text_config = LlamaConfig()
78
+
79
+ >>> # Initializing a default configuration
80
+ >>> configuration = UltravoxConfig(audio_config, text_config)
81
+
82
+ >>> # Initializing a completely untrained model from the configuration
83
+ >>> model = UltravoxForConditionalGeneration(configuration)
84
+
85
+ >>> # Accessing the model configuration
86
+ >>> configuration = model.config
87
+
88
+ >>> # Initialize a model from pretrained checkpoints and random projector weights
89
+ >>> config = UltravoxConfig(audio_model_id="facebook/wav2vec2-base-960h", text_model_id="meta-llama/Llama-2-7b-chat-hf")
90
+ ```"""
91
+
92
+ model_type = "ultravox"
93
+ is_composition = False
94
+
95
+ def __init__(
96
+ self,
97
+ audio_config: Optional[Dict[str, Any]] = None,
98
+ text_config: Optional[Dict[str, Any]] = None,
99
+ audio_model_id: Optional[str] = None,
100
+ text_model_id: Optional[str] = None,
101
+ ignore_index: int = -100,
102
+ hidden_size: int = 4096,
103
+ stack_factor: int = 8,
104
+ norm_init: float = 0.4,
105
+ projector_act: str = "swiglu",
106
+ text_model_lora_config: Optional[LoraConfigSimplified] = None,
107
+ audio_model_lora_config: Optional[LoraConfigSimplified] = None,
108
+ **kwargs,
109
+ ):
110
+ self.ignore_index = ignore_index
111
+
112
+ self.audio_model_id = audio_model_id
113
+ self.text_model_id = text_model_id
114
+
115
+ self.hidden_size = hidden_size
116
+ self.stack_factor = stack_factor
117
+ self.norm_init = norm_init
118
+ self.projector_act = projector_act
119
+
120
+ if text_model_id is not None:
121
+ self.text_config: transformers.LlamaConfig = (
122
+ transformers.AutoConfig.from_pretrained(text_model_id)
123
+ )
124
+ else:
125
+ text_config = text_config or {}
126
+ self.text_config = transformers.CONFIG_MAPPING[
127
+ text_config.get("model_type", "llama")
128
+ ](**text_config)
129
+
130
+ if audio_model_id is not None:
131
+ self.audio_config: transformers.PretrainedConfig = (
132
+ transformers.AutoConfig.from_pretrained(audio_model_id)
133
+ )
134
+ else:
135
+ audio_config = audio_config or {}
136
+ self.audio_config = transformers.CONFIG_MAPPING[
137
+ audio_config.get("model_type", "wav2vec2")
138
+ ](**audio_config)
139
+
140
+ self.text_model_lora_config = (
141
+ text_model_lora_config
142
+ if isinstance(text_model_lora_config, dict)
143
+ else dataclasses.asdict(text_model_lora_config or LoraConfigSimplified())
144
+ )
145
+ self.audio_model_lora_config = (
146
+ audio_model_lora_config
147
+ if isinstance(audio_model_lora_config, dict)
148
+ else dataclasses.asdict(audio_model_lora_config or LoraConfigSimplified())
149
+ )
150
+
151
+ self.vocab_size = self.text_config.vocab_size
152
+
153
+ self.initializer_range = self.text_config.initializer_range
154
+
155
+ super().__init__(**kwargs)
156
+
157
+ def to_diff_dict(self) -> Dict[str, Any]:
158
+ diff_dict = super().to_diff_dict()
159
+
160
+ # remove text_config and audio_config if text_model_id and audio_model_id are present
161
+ if self.text_model_id is not None:
162
+ diff_dict.pop("text_config", None)
163
+ if self.audio_model_id is not None:
164
+ diff_dict.pop("audio_config", None)
165
+
166
+ return diff_dict
ultravox_model.py ADDED
@@ -0,0 +1,660 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ from typing import Any, Dict, Optional, Set, Tuple, Union
3
+
4
+ import peft
5
+ import torch
6
+ import torch.nn as nn
7
+ import torch.nn.functional as F
8
+ import transformers
9
+ import transformers.activations
10
+ import transformers.modeling_outputs
11
+ import transformers.models
12
+ from transformers.models.whisper import modeling_whisper as whisper
13
+
14
+ # We must use relative import in this directory to allow uploading to HF Hub
15
+ # Even "from . import X" pattern doesn't work (undocumented and unclear why)
16
+ from .ultravox_config import LossConfig
17
+ from .ultravox_config import LossFunction
18
+ from .ultravox_config import UltravoxConfig
19
+
20
+
21
+ class UltravoxModel(transformers.LlamaPreTrainedModel):
22
+ """
23
+ The Ultravox model which consists of an audio encoder and a language model.
24
+
25
+ Audio input is processed by the audio encoder, then every `stack_factor` frames are stacked together and
26
+ projected to the language model's embedding space using a few linear layers.
27
+ The text is embedded by the language model as usual and then the audio and text embeddings are merged together.
28
+
29
+ A special token `<|audio|>` is used to indicate the start of the audio embeddings in the merged embeddings.
30
+
31
+ Parameters:
32
+ config: Model configuration class with all the parameters of the model.
33
+ """
34
+
35
+ config_class = UltravoxConfig
36
+ config: UltravoxConfig # for type hinting
37
+ # Usually we load encoder and LLM weights from a pretrained model separately, so they are allowed to be missing
38
+ _keys_to_ignore_on_load_missing = ["audio_tower.*", "language_model.*"]
39
+
40
+ def __init__(self, config: UltravoxConfig):
41
+ super().__init__(config)
42
+ self._register_load_state_dict_pre_hook(self._pre_load_state_dict_hook)
43
+
44
+ self.keep_params: Set[str] = set()
45
+ self.vocab_size = config.vocab_size
46
+
47
+ self.audio_tower = self._create_audio_tower(config)
48
+ self.multi_modal_projector = self._create_multi_modal_projector(config)
49
+ self.language_model = self._create_language_model(config)
50
+
51
+ # Determine no_split_modules dynamically to use with FSDP auto_wrap policy.
52
+ # FSDP throws an error if some of the layer types are not found in the model.
53
+ # This would be something like ["LlamaDecoderLayer", "WhisperEncoderLayer"]
54
+ self._no_split_modules = (self.language_model._no_split_modules or []) + (
55
+ self.audio_tower._no_split_modules or []
56
+ )
57
+
58
+ self.loss_config = LossConfig()
59
+ self.post_init()
60
+
61
+ def get_input_embeddings(self):
62
+ return self.language_model.get_input_embeddings()
63
+
64
+ def set_input_embeddings(self, value):
65
+ self.language_model.set_input_embeddings(value)
66
+
67
+ def get_output_embeddings(self):
68
+ return self.language_model.get_output_embeddings()
69
+
70
+ def set_output_embeddings(self, new_embeddings):
71
+ self.language_model.set_output_embeddings(new_embeddings)
72
+
73
+ def set_decoder(self, decoder):
74
+ self.language_model.set_decoder(decoder)
75
+
76
+ def get_decoder(self):
77
+ return self.language_model.get_decoder()
78
+
79
+ def tie_weights(self):
80
+ return self.language_model.tie_weights()
81
+
82
+ def set_loss_config(self, loss_config: LossConfig):
83
+ self.loss_config = loss_config
84
+
85
+ def _setup_cache(
86
+ self, cache_cls, max_batch_size: int, max_cache_len: Optional[int] = None
87
+ ):
88
+ self.language_model._setup_cache(cache_cls, max_batch_size, max_cache_len)
89
+
90
+ def _reorder_cache(self, past_key_values, beam_idx):
91
+ return self.language_model._reorder_cache(past_key_values, beam_idx)
92
+
93
+ def resize_token_embeddings(
94
+ self,
95
+ new_num_tokens: Optional[int] = None,
96
+ pad_to_multiple_of: Optional[int] = None,
97
+ ) -> nn.Embedding:
98
+ model_embeds = self.language_model.resize_token_embeddings(
99
+ new_num_tokens, pad_to_multiple_of
100
+ )
101
+ # update vocab size
102
+ self.config.text_config.vocab_size = model_embeds.num_embeddings
103
+ self.config.vocab_size = model_embeds.num_embeddings
104
+ self.vocab_size = model_embeds.num_embeddings
105
+ return model_embeds
106
+
107
+ def _compute_kl_loss(
108
+ self,
109
+ lm_output: transformers.modeling_outputs.CausalLMOutputWithPast,
110
+ labels: Optional[torch.Tensor] = None,
111
+ past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
112
+ alt_input_ids: Optional[torch.Tensor] = None,
113
+ alt_attention_mask: Optional[torch.Tensor] = None,
114
+ alt_labels: Optional[torch.Tensor] = None,
115
+ **kwargs,
116
+ ):
117
+ # disable gradient computation for the teacher model
118
+ with torch.no_grad():
119
+ # compute the teacher (text-only) model's distribution
120
+ alt_inputs_embeds = self.get_input_embeddings().forward(alt_input_ids)
121
+ alt_lm_output = self.language_model.forward(
122
+ inputs_embeds=alt_inputs_embeds,
123
+ labels=alt_labels,
124
+ attention_mask=alt_attention_mask,
125
+ past_key_values=past_key_values,
126
+ **kwargs,
127
+ )
128
+ # compute the KL divergence loss between the two models
129
+ kl_loss = F.kl_div(
130
+ F.log_softmax(
131
+ lm_output.logits[labels != -100] / self.loss_config.kl_temperature,
132
+ dim=-1,
133
+ ),
134
+ F.softmax(
135
+ alt_lm_output.logits[alt_labels != -100]
136
+ / self.loss_config.kl_temperature,
137
+ dim=-1,
138
+ ),
139
+ reduction="batchmean",
140
+ )
141
+ return {"loss": kl_loss}
142
+
143
+ def forward(
144
+ self,
145
+ input_ids: torch.Tensor,
146
+ audio_values: Optional[torch.FloatTensor] = None,
147
+ inputs_embeds: Optional[torch.FloatTensor] = None,
148
+ labels: Optional[torch.Tensor] = None,
149
+ attention_mask: Optional[torch.Tensor] = None,
150
+ audio_token_start_idx: Optional[torch.Tensor] = None,
151
+ audio_len: Optional[torch.Tensor] = None,
152
+ audio_token_len: Optional[torch.Tensor] = None,
153
+ past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
154
+ # the alt_* fields are needed for KL divergence loss
155
+ alt_input_ids: Optional[torch.Tensor] = None,
156
+ alt_attention_mask: Optional[torch.Tensor] = None,
157
+ alt_labels: Optional[torch.Tensor] = None,
158
+ **kwargs,
159
+ ) -> Union[Tuple, transformers.modeling_outputs.CausalLMOutputWithPast]:
160
+ """
161
+ Forward pass for the Ultravox model.
162
+
163
+ `input_ids` are the tokenized text input. They are embedded by the language model as usual.
164
+ `audio_values` are processed by the audio encoder and then every `stack_factor` frames are stacked together and
165
+ projected to the language model's embedding space using a few linear layers.
166
+ The audio and text embeddings are merged together. A special token `<|audio|>` is used to indicate the start
167
+ of the audio embeddings in the merged embeddings.
168
+
169
+ Args:
170
+ input_ids: The tokenized text input.
171
+ audio_values: The processed audio values.
172
+ inputs_embeds: The embeddings for the input tokens.
173
+ labels: The tokenized text labels.
174
+ attention_mask: The attention mask for the input.
175
+ position_ids: The position ids for the input.
176
+ past_key_values: The past key value cache for the language model attention layers.
177
+ **kwargs: Additional keyword arguments. Passed directly to the language model.
178
+ """
179
+ if inputs_embeds is None:
180
+ # B x T -> B x T x D
181
+ inputs_embeds = self.get_input_embeddings().forward(input_ids)
182
+
183
+ if audio_values is not None:
184
+ assert (
185
+ audio_token_start_idx is not None and audio_token_len is not None
186
+ ), "audio_token_start_idx and audio_token_len must be provided if audio_values are provided."
187
+ assert (
188
+ len(audio_token_start_idx) == len(audio_token_len) == len(audio_values)
189
+ ), "audio_token_start_idx, audio_token_len, and audio_values must have the same batch size."
190
+
191
+ # B x A/3200 x D
192
+ audio_tower_output = self.audio_tower.forward(
193
+ audio_values.to(self.audio_tower.dtype),
194
+ audio_len = audio_len
195
+ ).last_hidden_state
196
+ audio_tower_output = audio_tower_output.to(inputs_embeds.dtype)
197
+
198
+ audio_embeds = self.multi_modal_projector.forward(audio_tower_output)
199
+
200
+ # combine audio and text embeddings
201
+ for i, (audio, start, length) in enumerate(
202
+ zip(audio_embeds, audio_token_start_idx, audio_token_len)
203
+ ):
204
+ assert length <= audio.shape[0]
205
+ inputs_embeds[i, start : start + length].copy_(audio[:length])
206
+
207
+
208
+ lm_output = self.language_model.forward(
209
+ inputs_embeds=inputs_embeds,
210
+ labels=labels,
211
+ attention_mask=attention_mask,
212
+ past_key_values=past_key_values,
213
+ **kwargs,
214
+ )
215
+ if self.training:
216
+ if self.loss_config.loss_function == LossFunction.CrossEntropy:
217
+ return lm_output
218
+ elif self.loss_config.loss_function == LossFunction.KL_Divergence:
219
+ return self._compute_kl_loss(
220
+ lm_output=lm_output,
221
+ labels=labels,
222
+ past_key_values=past_key_values,
223
+ alt_input_ids=alt_input_ids,
224
+ alt_attention_mask=alt_attention_mask,
225
+ alt_labels=alt_labels,
226
+ **kwargs,
227
+ )
228
+ else:
229
+ raise ValueError(
230
+ f"Unsupported loss function: {self.loss_config.loss_function}"
231
+ )
232
+ else:
233
+ return lm_output
234
+
235
+ def prepare_inputs_for_generation(
236
+ self,
237
+ input_ids: torch.Tensor,
238
+ audio_values: Optional[torch.FloatTensor] = None,
239
+ audio_token_start_idx: Optional[torch.Tensor] = None,
240
+ audio_token_len: Optional[torch.Tensor] = None,
241
+ audio_len: Optional[torch.Tensor] = None,
242
+ past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
243
+ attention_mask: Optional[torch.Tensor] = None,
244
+ inputs_embeds: Optional[torch.Tensor] = None,
245
+ cache_position: Optional[torch.Tensor] = None,
246
+ **kwargs,
247
+ ) -> Dict[str, Any]:
248
+ model_input = self.language_model.prepare_inputs_for_generation(
249
+ input_ids=input_ids,
250
+ past_key_values=past_key_values,
251
+ attention_mask=attention_mask,
252
+ inputs_embeds=inputs_embeds,
253
+ cache_position=cache_position,
254
+ **kwargs,
255
+ )
256
+
257
+ # include audio information in model_input only when it is needed during prefilling
258
+ # audio_token_start_idx should always be relative to the current cache position
259
+ prefill_start_idx = 0 if cache_position is None else cache_position[0]
260
+ if (
261
+ audio_values is not None
262
+ and audio_token_start_idx is not None
263
+ and prefill_start_idx <= torch.max(audio_token_start_idx)
264
+ ):
265
+ model_input["audio_values"] = audio_values
266
+ model_input["audio_token_start_idx"] = (
267
+ audio_token_start_idx - prefill_start_idx
268
+ )
269
+ model_input["audio_token_len"] = audio_token_len
270
+ model_input["audio_len"] = audio_len
271
+
272
+ return model_input
273
+
274
+ @classmethod
275
+ def _create_multi_modal_projector(
276
+ cls, config: UltravoxConfig
277
+ ) -> "UltravoxProjector":
278
+ projector = UltravoxProjector(config)
279
+ projector.to(config.torch_dtype)
280
+ return projector
281
+
282
+ @classmethod
283
+ def _create_audio_tower(
284
+ cls, config: UltravoxConfig
285
+ ) -> Union[transformers.Wav2Vec2Model, "ModifiedWhisperEncoder"]:
286
+ if config.audio_model_id is not None:
287
+ if "whisper" in config.audio_model_id is not None:
288
+ audio_tower = ModifiedWhisperEncoder.from_pretrained(
289
+ config.audio_model_id, torch_dtype=config.torch_dtype
290
+ )
291
+ else:
292
+ audio_tower = transformers.AutoModel.from_pretrained(
293
+ config.audio_model_id, torch_dtype=config.torch_dtype
294
+ )
295
+ else:
296
+ if "whisper" in config.audio_config._name_or_path:
297
+ audio_tower = ModifiedWhisperEncoder(config.audio_config)
298
+ else:
299
+ with transformers.modeling_utils.no_init_weights():
300
+ # we only ever use from_config if the weights are retrained, hence initializing is not
301
+ # required. This makes the model quite creation faster since init on CPU is quite slow.
302
+ audio_tower = transformers.AutoModel.from_config(
303
+ config.audio_config
304
+ )
305
+
306
+ if isinstance(
307
+ audio_tower,
308
+ (transformers.Wav2Vec2BertModel, transformers.WhisperModel),
309
+ ):
310
+ # For these models we only need the encoder part
311
+ # Wav2Vec2BertModel -> Wav2Vec2BertEncoder
312
+ # WhisperModel -> WhisperEncoder
313
+ audio_tower = audio_tower.encoder
314
+
315
+ audio_tower = apply_lora(audio_tower, config.audio_model_lora_config)
316
+ return audio_tower
317
+
318
+ @classmethod
319
+ def _create_language_model(
320
+ cls, config: UltravoxConfig
321
+ ) -> transformers.LlamaForCausalLM:
322
+ if config.text_model_id is not None:
323
+ language_model = transformers.AutoModelForCausalLM.from_pretrained(
324
+ config.text_model_id,
325
+ attn_implementation=config._attn_implementation,
326
+ torch_dtype=config.torch_dtype,
327
+ )
328
+ else:
329
+ with transformers.modeling_utils.no_init_weights():
330
+ # we only ever use from_config if the weights are retrained, hence initializing is not
331
+ # required. This makes the model quite creation faster since init on CPU is quite slow.
332
+ language_model = transformers.AutoModelForCausalLM.from_config(
333
+ config.text_config,
334
+ attn_implementation=config._attn_implementation,
335
+ torch_dtype=config.torch_dtype,
336
+ )
337
+
338
+ language_model = apply_lora(language_model, config.text_model_lora_config)
339
+ return language_model
340
+
341
+ def merge_and_unload(self):
342
+ if isinstance(self.language_model, peft.PeftModel):
343
+ self.language_model = self.language_model.merge_and_unload()
344
+ # no need to download base language model weights anymore, so we can remove the id
345
+ self.config.text_model_id = None
346
+ self.keep_params.update(
347
+ set(
348
+ [
349
+ f"language_model.{name}"
350
+ for name, _ in self.language_model.named_parameters()
351
+ ]
352
+ )
353
+ )
354
+
355
+ if isinstance(self.audio_tower, peft.PeftModel):
356
+ self.audio_tower = self.audio_tower.merge_and_unload()
357
+ # no need to download base audio model weights anymore, so we can remove the id
358
+ self.config.audio_model_id = None
359
+ self.keep_params.update(
360
+ set(
361
+ [
362
+ f"audio_tower.{name}"
363
+ for name, _ in self.audio_tower.named_parameters()
364
+ ]
365
+ )
366
+ )
367
+
368
+ for param in ["text_model_lora_config", "audio_model_lora_config"]:
369
+ if hasattr(self.config, param):
370
+ delattr(self.config, param)
371
+
372
+ def push_to_hub(self, *args, **kwargs):
373
+ self.merge_and_unload()
374
+ self.to(self.language_model.dtype)
375
+ return super().push_to_hub(*args, **kwargs)
376
+
377
+ def save_pretrained(
378
+ self, *args, state_dict: Optional[Dict[str, Any]] = None, **kwargs
379
+ ):
380
+ if state_dict is None:
381
+ state_dict = super().state_dict()
382
+
383
+ named_params = dict(self.named_parameters())
384
+
385
+ state_dict = {
386
+ k: v
387
+ for k, v in state_dict.items()
388
+ if k in self.keep_params
389
+ or (k in named_params and named_params[k].requires_grad)
390
+ }
391
+
392
+ super().save_pretrained(*args, state_dict=state_dict, **kwargs)
393
+
394
+ def _pre_load_state_dict_hook(self, state_dict: Dict[str, Any], *args, **kwargs):
395
+ self.keep_params.update(set(state_dict.keys()))
396
+
397
+ def print_trainable_parameters(self):
398
+ """
399
+ Prints the number of trainable parameters in the model (reuses Peft model's method)
400
+ """
401
+ count_params = peft.peft_model.PeftModel.get_nb_trainable_parameters
402
+
403
+ trainable_params, all_param = count_params(self)
404
+
405
+ logging.info(
406
+ f"trainable params: {trainable_params:,d} || all params: {all_param:,d}"
407
+ f" || trainable%: {100 * trainable_params / all_param:.1f}%"
408
+ )
409
+
410
+ lm_trainable_params, lm_all_params = count_params(self.language_model)
411
+ audio_trainable_params, audio_all_params = count_params(self.audio_tower)
412
+
413
+ projector_trainable_params = (
414
+ trainable_params - lm_trainable_params - audio_trainable_params
415
+ )
416
+ projector_all_params = all_param - lm_all_params - audio_all_params
417
+
418
+ logging.info(
419
+ f"Trainable%: "
420
+ f" LLM: {100 * lm_trainable_params / lm_all_params:.1f}%"
421
+ f" || Audio Encoder: {100 * audio_trainable_params / audio_all_params:.1f}%"
422
+ f" || Projector: {100 * projector_trainable_params / projector_all_params:.1f}%"
423
+ )
424
+
425
+
426
+ def is_cache_empty(
427
+ past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]]
428
+ ) -> bool:
429
+ """
430
+ Check if the cache is empty.
431
+ """
432
+ if past_key_values is None:
433
+ return True
434
+ if isinstance(past_key_values, tuple):
435
+ return all(len(c) == 0 for c in past_key_values)
436
+ return past_key_values.get_seq_length() == 0
437
+
438
+
439
+ def apply_lora(model: torch.nn.Module, lora_config: dict) -> torch.nn.Module:
440
+ """
441
+ Applies LoRA finetuning to the model. If the `r` parameter is set to 0, the model is frozen instead.
442
+ """
443
+ lora_config = peft.LoraConfig(**lora_config or {})
444
+
445
+ if lora_config.r == 0:
446
+ # freeze the model entirely
447
+ for param in model.parameters():
448
+ param.requires_grad = False
449
+ else:
450
+ model = peft.get_peft_model(model, lora_config)
451
+
452
+ return model
453
+
454
+
455
+ class StackAudioFrames(nn.Module):
456
+ """
457
+ Stack the audio embedding frames to reduce the sequence length by a factor of `stack_factor`.
458
+
459
+ The number of output frames will be `ceil(T / stack_factor) + 1` where `T` is the number of input frames.
460
+ NOTE: the extra +1 is intentional: in case the number of audio tokens are over-estimated by the processor,
461
+ we want to make sure `processor.audio_token_replacement` (i.e. EOS) doesn't get leaked into the middle of embeddings.
462
+ In most cases this extra padding will get removed in the model's forward function so it has no effect.
463
+ """
464
+
465
+ def __init__(self, stack_factor: int = 8):
466
+ super().__init__()
467
+ self.stack_factor = stack_factor
468
+
469
+ def forward(self, audio_embeds: torch.Tensor) -> torch.Tensor:
470
+ B, T, C = audio_embeds.shape
471
+ T_pad = (T + self.stack_factor - 1) // self.stack_factor * self.stack_factor
472
+ audio_embeds = F.pad(audio_embeds, (0, 0, 0, T_pad - T + self.stack_factor))
473
+ B, T, C = audio_embeds.shape
474
+ audio_embeds = audio_embeds.view(
475
+ B, T // self.stack_factor, C * self.stack_factor
476
+ )
477
+ return audio_embeds
478
+
479
+
480
+ class RMSNorm(transformers.models.llama.modeling_llama.LlamaRMSNorm):
481
+ def __init__(self, hidden_size: int, init: float = 1, eps: float = 1e-6):
482
+ super().__init__(hidden_size=hidden_size, eps=eps)
483
+ self.weight.data.fill_(init)
484
+
485
+
486
+ class SwiGLU(nn.Module):
487
+ def forward(self, x):
488
+ x, gate = x.chunk(2, dim=-1)
489
+ return F.silu(gate) * x
490
+
491
+
492
+ class UltravoxProjector(nn.Sequential):
493
+ def __init__(self, config: UltravoxConfig):
494
+ super().__init__()
495
+ self.hidden_dim = config.hidden_size
496
+ self._pad_and_stack = StackAudioFrames(config.stack_factor)
497
+ dim = config.audio_config.hidden_size * config.stack_factor
498
+ self.ln_pre = RMSNorm(dim, init=config.norm_init)
499
+ self.linear_1 = nn.Linear(dim, self.hidden_dim, bias=False)
500
+ dim = self.hidden_dim
501
+ self.act = transformers.activations.get_activation(config.projector_act)
502
+ dim = dim // 2 if config.projector_act == "swiglu" else dim
503
+ self.linear_2 = nn.Linear(dim, config.text_config.hidden_size, bias=False)
504
+ self.ln_post = RMSNorm(config.text_config.hidden_size, init=config.norm_init)
505
+
506
+ def forward(self, audio_features: torch.Tensor) -> torch.Tensor:
507
+ audio_features = self._pad_and_stack(audio_features)
508
+ audio_features = self.ln_pre(audio_features)
509
+ hidden_states = self.linear_1(audio_features)
510
+ hidden_states = self.act(hidden_states)
511
+ hidden_states = self.linear_2(hidden_states)
512
+ hidden_states = self.ln_post(hidden_states)
513
+ return hidden_states
514
+
515
+
516
+ class ModifiedWhisperEncoder(whisper.WhisperEncoder, transformers.modeling_utils.ModuleUtilsMixin):
517
+ """
518
+ Encoder portion of OpenAI's Whisper model.
519
+
520
+ This implementation is a slightly modified version of HF Transformers' Whisper Encoder, with only a few fixes:
521
+ 1. base_model_prefix updated to allow for doing `.from_pretrained` directly on the encoder
522
+ 2. allow less than 30 second of audio padding to be passed in:
523
+ - relaxed ValueError check for `input_features` length to be less than or equal to `expected_seq_length` instead of strictly equal
524
+ - embed_pos is now sliced to match the length of `inputs_embeds`
525
+
526
+ Original: https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/modeling_whisper.py
527
+ """
528
+
529
+ base_model_prefix = "model.encoder"
530
+ _no_split_modules = ["WhisperEncoderLayer"]
531
+
532
+ def forward(
533
+ self,
534
+ input_features,
535
+ audio_len=None,
536
+ head_mask=None,
537
+ output_attentions=None,
538
+ output_hidden_states=None,
539
+ return_dict=None,
540
+ ):
541
+ expected_seq_length = (
542
+ self.config.max_source_positions
543
+ * self.conv1.stride[0]
544
+ * self.conv2.stride[0]
545
+ )
546
+ if input_features.shape[-1] > expected_seq_length:
547
+ raise ValueError(
548
+ f"Whisper expects the mel input features to be of length {expected_seq_length} or less, but found {input_features.shape[-1]}. Make sure to pad the input mel features to {expected_seq_length}."
549
+ )
550
+
551
+ output_attentions = (
552
+ output_attentions
553
+ if output_attentions is not None
554
+ else self.config.output_attentions
555
+ )
556
+ output_hidden_states = (
557
+ output_hidden_states
558
+ if output_hidden_states is not None
559
+ else self.config.output_hidden_states
560
+ )
561
+ return_dict = (
562
+ return_dict if return_dict is not None else self.config.use_return_dict
563
+ )
564
+ inputs_embeds = nn.functional.gelu(self.conv1(input_features))
565
+ inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
566
+
567
+ inputs_embeds = inputs_embeds.permute(0, 2, 1)
568
+ embed_pos = self.embed_positions.weight[: inputs_embeds.size(-2)]
569
+
570
+ hidden_states = inputs_embeds + embed_pos
571
+ hidden_states = nn.functional.dropout(
572
+ hidden_states, p=self.dropout, training=self.training
573
+ )
574
+
575
+ encoder_states = () if output_hidden_states else None
576
+ all_attentions = () if output_attentions else None
577
+
578
+ attention_mask = None
579
+ if audio_len != None:
580
+ audio_feature_len = self._get_feat_extract_output_lengths(audio_len)
581
+ batch_size = hidden_states.shape[0]
582
+ max_seq_len = hidden_states.shape[1]
583
+ attention_mask = (
584
+ torch.arange(max_seq_len, device=hidden_states.device)[None, :]
585
+ .expand(batch_size, -1)
586
+ .lt(audio_feature_len.view(batch_size, 1))
587
+ )
588
+ attention_mask = self.get_extended_attention_mask(
589
+ attention_mask,
590
+ None,
591
+ device=hidden_states.device,
592
+ dtype=hidden_states.dtype,
593
+ )
594
+
595
+ # check if head_mask has a correct number of layers specified if desired
596
+ if head_mask is not None:
597
+ assert head_mask.size()[0] == (
598
+ len(self.layers)
599
+ ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
600
+
601
+ for idx, encoder_layer in enumerate(self.layers):
602
+ if output_hidden_states:
603
+ encoder_states = encoder_states + (hidden_states,)
604
+ # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
605
+ to_drop = False
606
+ if self.training:
607
+ dropout_probability = torch.rand([])
608
+ if dropout_probability < self.layerdrop: # skip the layer
609
+ to_drop = True
610
+
611
+ if to_drop:
612
+ layer_outputs = (None, None)
613
+ else:
614
+ if self.gradient_checkpointing and self.training:
615
+ layer_outputs = self._gradient_checkpointing_func(
616
+ encoder_layer.__call__,
617
+ hidden_states,
618
+ attention_mask,
619
+ (head_mask[idx] if head_mask is not None else None),
620
+ output_attentions,
621
+ )
622
+ else:
623
+ layer_outputs = encoder_layer(
624
+ hidden_states,
625
+ attention_mask,
626
+ layer_head_mask=(
627
+ head_mask[idx] if head_mask is not None else None
628
+ ),
629
+ output_attentions=output_attentions,
630
+ )
631
+
632
+ hidden_states = layer_outputs[0]
633
+
634
+ if output_attentions:
635
+ all_attentions = all_attentions + (layer_outputs[1],)
636
+
637
+ hidden_states = self.layer_norm(hidden_states)
638
+ if output_hidden_states:
639
+ encoder_states = encoder_states + (hidden_states,)
640
+
641
+ if not return_dict:
642
+ return tuple(
643
+ v
644
+ for v in [hidden_states, encoder_states, all_attentions]
645
+ if v is not None
646
+ )
647
+ return transformers.modeling_outputs.BaseModelOutput(
648
+ last_hidden_state=hidden_states,
649
+ hidden_states=encoder_states,
650
+ attentions=all_attentions,
651
+ )
652
+
653
+
654
+ UltravoxConfig.register_for_auto_class()
655
+ UltravoxModel.register_for_auto_class()
656
+
657
+ transformers.AutoConfig.register("ultravox", UltravoxConfig)
658
+ transformers.AutoModel.register(UltravoxConfig, UltravoxModel)
659
+
660
+ transformers.activations.ACT2FN["swiglu"] = SwiGLU
ultravox_pipeline.py ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ from typing import Any, Dict, List, Optional
3
+
4
+ import numpy as np
5
+ import transformers
6
+
7
+ # We must use relative import in this directory to allow uploading to HF Hub
8
+ # Even "from . import X" pattern doesn't work (undocumented and unclear why)
9
+ from .ultravox_model import UltravoxModel
10
+ from .ultravox_processing import UltravoxProcessor
11
+
12
+
13
+ class UltravoxPipeline(transformers.Pipeline):
14
+ def __init__(
15
+ self,
16
+ model: UltravoxModel,
17
+ tokenizer: Optional[transformers.PreTrainedTokenizerBase] = None,
18
+ audio_processor: Optional[transformers.ProcessorMixin] = None,
19
+ **kwargs
20
+ ):
21
+ if tokenizer is None:
22
+ try:
23
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
24
+ model.config._name_or_path
25
+ )
26
+ except:
27
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
28
+ model.config.text_model_id or model.config.text_config._name_or_path
29
+ )
30
+
31
+ if audio_processor is None:
32
+ audio_processor = transformers.AutoProcessor.from_pretrained(
33
+ model.config.audio_model_id or model.config.audio_config._name_or_path
34
+ )
35
+
36
+ super().__init__(model=model, tokenizer=tokenizer, **kwargs)
37
+
38
+ self.processor = UltravoxProcessor(
39
+ audio_processor=audio_processor,
40
+ tokenizer=tokenizer,
41
+ stack_factor=model.config.stack_factor,
42
+ )
43
+
44
+ def _sanitize_parameters(self, **kwargs):
45
+ generation_keys = ["temperature", "max_new_tokens", "repetition_penalty"]
46
+ generation_kwargs = {k: kwargs[k] for k in kwargs if k in generation_keys}
47
+ return {}, generation_kwargs, {}
48
+
49
+ def preprocess(self, inputs: Dict[str, Any]):
50
+ turns: list = inputs.get("turns", [])
51
+
52
+ audio = inputs.get("audio", None)
53
+ # Convert to float32 if needed.
54
+ if isinstance(audio, np.ndarray):
55
+ if audio.dtype == np.float64:
56
+ audio = audio.astype(np.float32)
57
+ elif audio.dtype == np.int16:
58
+ audio = audio.astype(np.float32) / np.float32(32768.0)
59
+ elif audio.dtype == np.int32:
60
+ audio = audio.astype(np.float32) / np.float32(2147483648.0)
61
+
62
+ if audio is not None and (len(turns) == 0 or turns[-1]["role"] != "user"):
63
+ prompt = inputs.get("prompt", "<|audio|>")
64
+ if "<|audio|>" not in prompt:
65
+ logging.warning(
66
+ "Prompt does not contain '<|audio|>', appending '<|audio|>' to the end of the prompt."
67
+ )
68
+
69
+ prompt += " <|audio|>"
70
+ turns.append({"role": "user", "content": prompt})
71
+
72
+ text = self.processor.tokenizer.apply_chat_template(
73
+ turns, add_generation_prompt=True, tokenize=False
74
+ )
75
+
76
+ if "sampling_rate" not in inputs and audio is not None:
77
+ logging.warning(
78
+ "No sampling rate provided, using default of 16kHz. We highly recommend providing the correct sampling rate."
79
+ )
80
+
81
+ output = self.processor(
82
+ text=text,
83
+ audio=audio,
84
+ sampling_rate=inputs.get("sampling_rate", 16000),
85
+ )
86
+ if "audio_values" in output:
87
+ output["audio_values"] = output["audio_values"].to(self.model.dtype)
88
+
89
+ return output
90
+
91
+ def _forward(
92
+ self,
93
+ model_inputs: Dict[str, Any],
94
+ temperature: Optional[float] = None,
95
+ max_new_tokens: Optional[int] = None,
96
+ repetition_penalty: float = 1.1,
97
+ ) -> List[int]:
98
+ temperature = temperature or None
99
+ do_sample = temperature is not None
100
+
101
+ terminators = [self.tokenizer.eos_token_id]
102
+ if "<|eot_id|>" in self.tokenizer.added_tokens_encoder:
103
+ terminators.append(self.tokenizer.convert_tokens_to_ids("<|eot_id|>"))
104
+
105
+ input_len = model_inputs["input_ids"].shape[1]
106
+
107
+ outputs = self.model.generate(
108
+ **model_inputs,
109
+ do_sample=do_sample,
110
+ temperature=temperature,
111
+ max_new_tokens=max_new_tokens,
112
+ repetition_penalty=repetition_penalty,
113
+ eos_token_id=terminators
114
+ )
115
+ return outputs[0][input_len:]
116
+
117
+ def postprocess(self, model_outputs) -> str:
118
+ output_text = self.tokenizer.decode(model_outputs, skip_special_tokens=True)
119
+ return output_text
120
+
121
+
122
+ transformers.pipelines.PIPELINE_REGISTRY.register_pipeline(
123
+ "ultravox-pipeline",
124
+ pipeline_class=UltravoxPipeline,
125
+ pt_model=transformers.AutoModel,
126
+ type="multimodal",
127
+ )
ultravox_processing.py ADDED
@@ -0,0 +1,212 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Union
2
+
3
+ import numpy as np
4
+ import torch
5
+ import transformers
6
+
7
+ from .ultravox_config import UltravoxConfig
8
+
9
+
10
+ class UltravoxProcessor(transformers.ProcessorMixin):
11
+ """
12
+ Constructs an Ultravox processor which wraps an audio processor and a tokenizer into a single processor.
13
+ Args:
14
+ audio_processor: The audio processor for the audio encoder.
15
+ tokenizer: The tokenizer for the language model.
16
+ """
17
+
18
+ attributes = ["audio_processor", "tokenizer"]
19
+ audio_processor_class = (
20
+ "Wav2Vec2Processor",
21
+ "SeamlessM4TFeatureExtractor",
22
+ "WhisperProcessor",
23
+ "Wav2Vec2BertProcessor",
24
+ )
25
+ tokenizer_class = (
26
+ "PreTrainedTokenizer",
27
+ "PreTrainedTokenizerFast",
28
+ )
29
+
30
+ tokenizer: transformers.PreTrainedTokenizerBase
31
+ audio_processor: transformers.ProcessorMixin
32
+
33
+ def __init__(
34
+ self,
35
+ audio_processor=None,
36
+ tokenizer=None,
37
+ audio_padding: str = "longest",
38
+ encoder_ds_factor: int = 320,
39
+ stack_factor: int = 8,
40
+ audio_placeholder: str = "<|audio|>",
41
+ ):
42
+ """
43
+ Args:
44
+ audio_processor: The audio processor for the audio encoder.
45
+ tokenizer: The tokenizer for the language model.
46
+ audio_padding: The padding strategy for the audio encoder.
47
+ encoder_ds_factor: The downsample factor of the audio encoder.
48
+ stack_factor: The factor by which the audio encoder output is stacked in the multimodal projector.
49
+ audio_placeholder: The placeholder for the audio in the text.
50
+ """
51
+ self.audio_padding = audio_padding
52
+ self.encoder_ds_factor = encoder_ds_factor
53
+ self.stack_factor = stack_factor
54
+ self.audio_placeholder = audio_placeholder
55
+ self.audio_token_replacement = tokenizer.eos_token
56
+ assert (
57
+ self.audio_token_replacement is not None
58
+ ), "The tokenizer has no EOS token. Cannot recover."
59
+ if tokenizer.pad_token_id is None:
60
+ tokenizer.pad_token_id = tokenizer.eos_token_id
61
+
62
+ super().__init__(audio_processor=audio_processor, tokenizer=tokenizer)
63
+
64
+ @classmethod
65
+ def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
66
+ config: UltravoxConfig = transformers.AutoConfig.from_pretrained(
67
+ pretrained_model_name_or_path, **kwargs
68
+ )
69
+ audio_processor = transformers.AutoProcessor.from_pretrained(
70
+ config.audio_model_id
71
+ or config.audio_config._name_or_path
72
+ or "facebook/wav2vec2-base-960h"
73
+ )
74
+
75
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
76
+ pretrained_model_name_or_path, **kwargs
77
+ )
78
+ tokenizer.padding_side = "left"
79
+ tokenizer.pad_token = tokenizer.eos_token
80
+
81
+ return cls(
82
+ audio_processor=audio_processor,
83
+ tokenizer=tokenizer,
84
+ stack_factor=config.stack_factor,
85
+ )
86
+
87
+ def __call__(
88
+ self,
89
+ text: Optional[str] = None,
90
+ audio: Optional[Union[np.ndarray, torch.Tensor]] = None,
91
+ sampling_rate: Optional[int] = None,
92
+ return_tensors: Optional[
93
+ Union[str, transformers.TensorType]
94
+ ] = transformers.TensorType.PYTORCH,
95
+ **kwargs,
96
+ ) -> transformers.BatchFeature:
97
+ """
98
+ Main method to prepare for the model one text sequence and audio. This method forwards the `text`
99
+ and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode
100
+ the text. To prepare the audio(s), this method forwards the `audio`, `sampling_rate` and `kwargs` arguments to
101
+ audio processor's [`~Wav2Vec2Processor.__call__`] if `audio` is not `None`. Please refer to the docstring
102
+ of the above two methods for more information.
103
+ Args:
104
+ text (`str`, `List[str]`):
105
+ The sequence to be encoded. Sequence can be a string or (pretokenized string).
106
+ audio (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
107
+ The audio to be prepared. Audio can be NumPy array or PyTorch tensor. In case of a
108
+ NumPy array/PyTorch tensor, each audio should be of shape (C, T), where C is a number of channels, and T the
109
+ sample length of the audio.
110
+ sampling_rate (`int`, *optional*, defaults to 16000):
111
+ Sampling rate of the input audio. We expect 16kHz audio. Don't change this value unless you know what
112
+ you are doing.
113
+ return_tensors (`str` or [`~utils.TensorType`], *optional*):
114
+ If set, will return tensors of a particular framework. Acceptable values are:
115
+ - `'tf'`: Return TensorFlow `tf.constant` objects.
116
+ - `'pt'`: Return PyTorch `torch.Tensor` objects.
117
+ - `'np'`: Return NumPy `np.ndarray` objects.
118
+ - `'jax'`: Return JAX `jnp.ndarray` objects.
119
+ Returns:
120
+ [`BatchFeature`]: A [`BatchFeature`] with the following fields:
121
+ - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
122
+ - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
123
+ `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
124
+ `None`).
125
+ - **audio_values** -- Processed audio values to be fed to a model. Returned when `audio` is not `None`.
126
+ - **audio_token_len** -- Predicted number of audio frames: this value is guaranteed to be a close upper bound.
127
+ Returned when `audio` is not `None`.
128
+ - **audio_token_start_idx** -- The index in the tokenized text where the audio starts. Returned when `audio` is not `None`.
129
+ """
130
+ # TODO: Add support for multiple audio and text inputs.
131
+ data = {}
132
+ audio_embed_frames = 0
133
+ if audio is not None and len(audio) > 0:
134
+ if self.audio_padding == "max_length":
135
+ # 30 seconds is the expected length for Whisper
136
+ assert sampling_rate is not None, "Sampling rate must be provided."
137
+ audio_len = 30 * sampling_rate
138
+ else:
139
+ audio_len = max([a.shape[-1] for a in audio])
140
+ # It's guaranteed that the number of frames is less than or equal to this amount.
141
+ # For Whisper this is exact AFAICT, but for Wav2Vec2 it's an upper bound.
142
+ # Currently, StackAudioFrames makes sure an over-estimation won't cause issues by padding the audio embeddings.
143
+ nb_encoder_frames = int(round(audio_len / self.encoder_ds_factor + 1e-4))
144
+ audio_embed_frames = int(np.ceil(nb_encoder_frames / self.stack_factor))
145
+ data["audio_token_len"] = [audio_embed_frames]
146
+
147
+ # Main audio processing. The processor is model-specific.
148
+ x = self.audio_processor(
149
+ audio,
150
+ sampling_rate=sampling_rate,
151
+ padding="longest",
152
+ max_length=audio_len,
153
+ return_attention_mask=True,
154
+ **kwargs,
155
+ )
156
+ if "input_features" in x:
157
+ data["audio_values"] = x.input_features
158
+ else:
159
+ data["audio_values"] = x.input_values
160
+ if self.audio_padding == "max_length":
161
+ data["audio_len"] = x.attention_mask.sum(-1) - 1
162
+ else:
163
+ data["audio_len"] = [data["audio_values"].shape[-1]]
164
+
165
+ if text is not None:
166
+ assert isinstance(
167
+ text, str
168
+ ), "Text must be a string. Batch mode not supported yet."
169
+ if self.audio_placeholder in text:
170
+ if "audio_token_len" not in data:
171
+ raise ValueError(
172
+ f"audio must be provided when using audio placeholder ({self.audio_placeholder}) in text."
173
+ )
174
+
175
+ start_idx = len(
176
+ self.tokenizer.encode(
177
+ text[: text.index(self.audio_placeholder)],
178
+ add_special_tokens=False,
179
+ )
180
+ )
181
+ data["audio_token_start_idx"] = [start_idx]
182
+
183
+ # Replace the audio placeholder with the audio token.
184
+ # e.g. "Transcribe\n<|audio|>" -> "Transcribe </s></s></s></s></s></s></s></s>"
185
+ # where the number of </s> is the number of audio frames.
186
+ text = text.replace(
187
+ self.audio_placeholder,
188
+ self.audio_token_replacement * audio_embed_frames,
189
+ )
190
+
191
+ # Special tokens like BOS should already have been added by the caller.
192
+ data.update(self.tokenizer([text], add_special_tokens=False, **kwargs))
193
+
194
+ return transformers.BatchFeature(data=data, tensor_type=return_tensors)
195
+
196
+ def batch_decode(self, *args, **kwargs):
197
+ return self.tokenizer.batch_decode(*args, **kwargs)
198
+
199
+ def decode(self, *args, **kwargs):
200
+ return self.tokenizer.decode(*args, **kwargs)
201
+
202
+ @property
203
+ def model_input_names(self):
204
+ tokenizer_input_names = self.tokenizer.model_input_names
205
+ audio_processor_input_names = self.audio_processor.model_input_names
206
+ return list(set(tokenizer_input_names + audio_processor_input_names))
207
+
208
+
209
+ UltravoxProcessor.register_for_auto_class()
210
+
211
+ transformers.AutoProcessor.register(UltravoxConfig, UltravoxProcessor)
212
+