Support for custom SentenceTransformer models

[wwymak]

Model description

I have a custom SentenceTransformer model that is a custom class (And also quite nested), so on the top level the modules.json file look like

[
  {
    "idx": 0,
    "name": "0",
    "path": "0_MyModelClass",
    "type": "custom_models.MyModelClass"
  }
]

This loads correctly if I use SentenceTransformers directly, but when loading in infinity it complains that the .auto_model attribute is missing (thrown by these lines https://github.com/michaelfeil/infinity/blob/main/libs/infinity_emb/infinity_emb/transformer/embedder/sentence_transformer.py#L81-L93 ). If this sort of custom model can be supported, or if you can give me some guidance on the correct way to save the model, that would be great.

Open source status & huggingface transformers.

• The model implementation is available on transformers
• The model weights are available on huggingface-hub
• I verified that the model is currently not running in the latest version pip install infinity_emb[all] --upgrade
• I made the authors of the model aware that I want to use it with infinity_emb & check if they are aware of the issue.

[michaelfeil]

This is not supposed to work, name needs to be "auto_model". I would recommend exporting it to onnx (using your favorite toolchain)

[slobstone]

I've seen a few cases with sentence transformers where teams have used slightly bespoke structures (wrappers around a single Transformer at the heart of it all) -- this may get even more prevalent with Release 3.1.0 . Tweaking this line to "something along the lines of" fm = [x for x in self.modules() if hasattr(x, "auto_model")][0] would work in cases where there is a single Transformer. (Also, you may have to turn off the warmup (--no-model-warmup) depending on whether or not your model expects a particular form of input)

[michaelfeil]

Thanks. Perhjaps `hasattr(fm, "auto_model") would be helpful. be prepared that there is little / no optimizations doable for a generic CustomModel