Feature : Support SPLADE pooling

[tsensei]

Feature request

I want to run the https://huggingface.co/prithivida/Splade_PP_en_v1 using infinity. While it loads the model, but the output isn't the sparse representation. It being a sparse embedding model, the huggingface Text Embedding Interface documentation says we need to set the pooling to splade. I am not sure if that solves the issue and will output sparse representation but the idea is to support sparse representation output.

Motivation

Embedding support for hybrid retrieval, infinity already support Dense models, supporting Sparse model would make it more versatile and drop in tool for hybrid retrieval approach

Your contribution

Went through previous issues but found nothing relevant

[tsensei]

Or this model as well opensearch-project/opensearch-neural-sparse-encoding-doc-v2-distill

[tsensei]

@michaelfeil pardon me for mentioning here but is there any possibility of this getting implemented anytime soon? Otherwise I'll have to strategize differently

Thanks for the great project btw