ONNX-GPU

[michaelfeil]

This PR aims to add optional GPU support, only if onnxruntime-gpu is installed. Reference: michaelfeil/infinity#26

Closes #52

[equal-l2]

You might want to add the extras section to the pyproject.toml, to indicate that fastembed can use onnxruntime-gpu as an optional dependency.
For details about extras section in pyproject.toml, see https://python-poetry.org/docs/pyproject/#extras.

[michaelfeil]

Yes, I have done this in my personal package. Note that I stopped contributing here, as none of the PRs was getting attention.

[logan-markewich]

would be cool if this got merged...

[generall]

I have the following concerns before merging this PR:

• Does it introduce new cuda-related dependencies into the package?
  • If not, how can it work with cuda?
  • Does it requires manual installation of dependencies? If so, this should be documented.
  • If it introduces new dependencies, is it possible to make them optional? Usually cuda-related packages are very big and I would prefer to avoid them if I am working in CPU-only environment
• If I have GPU on my machine, how can I force it to use CPU-only? This should be documented

[Anexen]

I believe it would be beneficial to allow users to specify a list of providers. The FlagEmbeddings class can be enhanced to accept an additional parameter called onnx_providers.

My current workaround is as follows:

embedding = FlagEmbedding()
embedding.model.model.set_providers(
    ["CUDAExecutionProvider", "CPUExecutionProvider"]
)

[pazoff]

It would be very, very cool if this got merged...

[valentimarco]

would be very cool if this got merged!

[NirantK]

Closing this in favour of #214, closes #52 and #97