axinc-ai / ailia-models

The collection of pre-trained, state-of-the-art AI models for ailia SDK
2k stars 318 forks source link

Implement sdxl-turbo #1469

Closed ooe1123 closed 1 month ago

ooe1123 commented 4 months ago

https://github.com/axinc-ai/ailia-models/issues/1435 のPRです。

kyakuno commented 1 month ago

モデルアップロード済み。 https://storage.googleapis.com/ailia-models/sdxl-turbo/text_encoder.onnx など

kyakuno commented 1 month ago

tokenizerとtokenizer_2のdiff。

kyakuno@mbakk sdxl-turbo % diff tokenizer/merges.txt tokenizer_2/merges.txt 
kyakuno@mbakk sdxl-turbo % diff tokenizer/vocab.json tokenizer_2/vocab.json
kyakuno@mbakk sdxl-turbo % diff tokenizer/tokenizer_config.json tokenizer_2/tokenizer_config.json
3a4,11
>     "0": {
>       "content": "!",
>       "lstrip": false,
>       "normalized": false,
>       "rstrip": false,
>       "single_word": false,
>       "special": true
>     },
27c35
<   "pad_token": "<|endoftext|>",
---
>   "pad_token": "!",
kyakuno@mbakk sdxl-turbo % diff tokenizer/special_tokens_map.json tokenizer_2/special_tokens_map.json
17c17
<     "content": "<|endoftext|>",
---
>     "content": "!",
19c19
<     "normalized": true,
---
>     "normalized": false,
kyakuno commented 1 month ago

tokenizer_2は、padトークンとして0が追加されている。vocabは同じ。

kyakuno commented 1 month ago

clipのデフォルトは  target._pad_token_id = 49407 なので、  tokenizer.add_special_tokens({'pad_token': '!'}) を呼べば等価になる。