wsl-oracle 安装 omlutils

1. 安装 cmake 和 gcc-c++

sudo dnf install -y cmake gcc-c++

2. 安装 omlutils

pip install omlutils-0.10.0-cp312-cp312-linux_x86_64.whl

不需要安装 requirements.txt，特别是里面有torch==2.2.0+cpu，会卸载掉支持 GPU 的 torch。

--extra-index-url https://download.pytorch.org/whl/cpu
torch==2.2.0+cpu

3. 使用 omlutils 创建 onnx 模型

安装 sentencepiece，

pip install sentencepiece

修复omlutils部分代码使其支持支持大于1GB，是tokenizer是XLMRobertaTokenizer的模型。

vi /home/oracle/miniconda/envs/learn-oracle23c/lib/python3.12/site-packages/omlutils/_pipeline/steps.py--- beforesize_threshold = quant_limit if is_quantized else 0.99e9
------ aftersize_threshold = quant_limit if is_quantized else 0.99e9 * 5
--- after

vi /home/oracle/miniconda/envs/learn-oracle23c/lib/python3.12/site-packages/omlutils/_pipeline/steps.py--- beforedef validateBertTokenizer(self,tokenizer):supportedTokenizer=[transformers.models.bert.BertTokenizer, transformers.models.distilbert.DistilBertTokenizer,transformers.models.mpnet.MPNetTokenizer]cls=tokenizer.__class__if(cls not in supportedTokenizer):raise ValueError(f"Unsupported tokenizer {cls}")
------ afterdef validateBertTokenizer(self,tokenizer):supportedTokenizer=[transformers.models.bert.BertTokenizer, transformers.models.distilbert.DistilBertTokenizer,transformers.models.mpnet.MPNetTokenizer,transformers.models.xlm_roberta.tokenization_xlm_roberta.XLMRobertaTokenizer]cls=tokenizer.__class__if(cls not in supportedTokenizer):raise ValueError(f"Unsupported tokenizer {cls}")
--- after

vi /home/oracle/miniconda/envs/learn-oracle23c/lib/python3.12/site-packages/omlutils/_onnx_export/tokenizer_export.py--- before
TOKENIZER_MAPPING = {transformers.models.bert.BertTokenizer: SupportedTokenizers.BERT,transformers.models.clip.CLIPTokenizer: SupportedTokenizers.CLIP,transformers.models.distilbert.DistilBertTokenizer: SupportedTokenizers.BERT,transformers.models.gpt2.GPT2Tokenizer: SupportedTokenizers.GPT2,#transformers.models.llama.LlamaTokenizer: SupportedTokenizers.SENTENCEPIECE,# transformers.models.mluke.MLukeTokenizer: SupportedTokenizers.SENTENCEPIECE,transformers.models.mpnet.MPNetTokenizer: SupportedTokenizers.BERT,# transformers.models.roberta.tokenization_roberta.RobertaTokenizer: SupportedTokenizers.ROBERTA,# transformers.models.xlm_roberta.XLMRobertaTokenizer: SupportedTokenizers.SENTENCEPIECE,
}
------ after
TOKENIZER_MAPPING = {transformers.models.bert.BertTokenizer: SupportedTokenizers.BERT,transformers.models.clip.CLIPTokenizer: SupportedTokenizers.CLIP,transformers.models.distilbert.DistilBertTokenizer: SupportedTokenizers.BERT,transformers.models.gpt2.GPT2Tokenizer: SupportedTokenizers.GPT2,#transformers.models.llama.LlamaTokenizer: SupportedTokenizers.SENTENCEPIECE,# transformers.models.mluke.MLukeTokenizer: SupportedTokenizers.SENTENCEPIECE,transformers.models.mpnet.MPNetTokenizer: SupportedTokenizers.BERT,# transformers.models.roberta.tokenization_roberta.RobertaTokenizer: SupportedTokenizers.ROBERTA,transformers.models.xlm_roberta.XLMRobertaTokenizer: SupportedTokenizers.SENTENCEPIECE,
}
---

创建 multilingual_e5_small.py，内容如下，

from omlutils import EmbeddingModel, EmbeddingModelConfig
print(f"start...")
config = EmbeddingModelConfig.from_template("text", max_seq_length=512)
em = EmbeddingModel(model_name="intfloat/multilingual-e5-small", config=config)
em.export2file("multilingual_e5_small", output_dir=".")
print(f"complete...")

创建 onnx 模型，

python multilingual_e5_small.py

程序执行完成后，会创建一个 multilingual_e5_small.onnx 文件。

（可选）升级transformers，

pip install -U transformers

完结！