[Bug]: Fastembed's TextEmbedding does not cache locally when `specific_model_path` is provided

[Tayomide]

What happened?

TextEmbedding(
  model_name=text_model_name,
  specific_model_path="/specific/local/url",
  cache_dir = "/cache"
)

creates the cache directory but does not create the snapshots, blobs, and refs that are usually present when downloading from hugging face

My current /specific/local/url/ contains these files: config.json, model.onnx, model_optimized.onnx, special_tokens_map.json, tokenizer.json, tokenizer_config.json

Could there be some missing files on my end that are needed for the cache logic to work?

What is the expected behaviour?

A cache directory with model file and other metadata that is interactive with fastembed's TextEmbedding API

A minimal reproducible example

from fastembed import TextEmbedding

emb = TextEmbedding(
  model_name=text_model_name,
  specific_model_path="/specific/local/url",
  cache_dir = "/cache"
)

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

Python 3.11.9

FastEmbed version

v0.6.0

What os are you seeing the problem on?

Linux

Relevant stack traces and/or logs

[joein]

specific_model_path was introduced exactly to avoid creating a structure with snapshots, blobs, refs, etc
cache directory, since it is created by huggingface modules and sometimes you can't download models from huggingface since some security procedures are required.

The assumption is when you provide specific_model_path - you already have everything on disk as plain files and want to load your model into memory without any interaction with huggingface modules.

Cache directory might still be created, but should remain empty

[Tayomide]

Thank you for your reply!

That makes perfect sense

Currently my team mounts an external storage to our docker container where the fastembed logic is called. Using specific_model_path on the mounted path is a bit faster and reliable in retrieving the model but still slower than a local cache. Would it be possible to have a toggle in situations like this to cache the model after the first call?

[joein]

How slower is it?

There is no functionality which will allow to move files from specific_model_path to a cache directory, and it is unlikely that we will implement it.

Could you maybe either copy/download the model during building the container or just copy files after container startup with docker cp ?

[Tayomide]

That works!

I will send the update back and that should resolve our current problem

Also it is not that much slower, but it is an external storage and we plan on keeping the models as close to the code as we can

Thank you!

[Tayomide]

Thank you so much!

This issue has been resolved