[Bug]: Model no longer supported in 0.7.3

[constance-scherer]

What happened?

A sentence transformers model from HuggingFace is no longer supported in v0.7.3, while it was supported in v0.7.2.
The model is sentence-transformers/paraphrase-xlm-r-multilingual-v1. To my knowledge, this specific model is not listed in the supported models list, but nonetheless, it works with v0.7.2.

What is the expected behaviour?

Same behaviour as in v0.7.2, which is support for this model.

A minimal reproducible example

from fastembed import TextEmbedding

model = TextEmbedding(model_name="sentence-transformers/paraphrase-xlm-r-multilingual-v1")

What Python version are you on? e.g. python --version

3.12

FastEmbed version

v0.7.3

What os are you seeing the problem on?

Linux

Relevant stack traces and/or logs

ValueError: Model sentence-transformers/paraphrase-xlm-r-multilingual-v1 is not supported in TextEmbedding. Please check the supported models using `TextEmbedding.list_supported_models()`

[coderabbitai]

A summary of the changes CodeRabbit can apply:

• Add a new entry for sentence-transformers/paraphrase-xlm-r-multilingual-v1 to the supported_pooled_normalized_models list in fastembed/text/pooled_normalized_embedding.py, including metadata (dimensions: 768, description, license: apache-2.0, source: HuggingFace sentence-transformers/paraphrase-xlm-r-multilingual-v1, model_file: onnx/model.onnx) and configure it for mean pooling with L2 normalization to restore model support.

• Update fastembed/text/pooled_normalized_embedding.py by adding a new DenseModelDescription entry for "sentence-transformers/paraphrase-xlm-r-multilingual-v1" (dim=768, description, apache-2.0 license, size_in_GB=1.00, sources pointing to HF, model_file "onnx/model.onnx").

• ✅ Create PR with these edits
• 📋 Get copyable edits

[constance-scherer]

Update: after further investigation, it seems that this model doesn't actually work when in a brand new environment. However, there might be some interactions with a HuggingFace cache somehow, somewhere, because the loading of this model through a TextEmbedding has worked in the past, without the use of add_custom_model

[joein]

Hey @constance-scherer

It is really strange that it worked in fastembed 0.7.2, because we have never supported it.
Could it be that you actually used sentence-transformers instead of fastembed?

TextEmbedding could not (and can't) be instantiated with an unsupported model