[Feature] Serving embedding and reranking model using vLLM

[lvliang-intel]

Priority

P1-Stopper

OS type

Ubuntu

Hardware type

Xeon-GNR

Running nodes

Single Node

Description

Feature: Serving Embedding and Reranking Models Using vLLM on Xeon and Gaudi
Description:
Integrate vLLM as a serving framework to enhance the performance and scalability of embedding and reranking models. This feature involves:

Leveraging vLLM's high-throughput serving capabilities to efficiently handle embedding and reranking requests.
Integration with the ChatQnA pipeline.
Optimizing the vLLM configuration for use cases involving embeddings and reranking, ensuring lower latency and better resource utilization.
Comparing vLLM's performance against the current TEI to determine the best setup for production.

Expected Outcome:

Applied another serving framework for embedding and reranking models, expect better performance on Gaudi.
Improved throughput for embedding and reranking services.
Enhanced flexibility to switch between serving frameworks based on specific requirements.

[lvliang-intel]

The microservice itself will be done by GenAIComps feature opea-project/GenAIComps#956. And if the embedding/reranking performance serving by vLLM is better than TGI, we will update the ChatQnA example.

[eero-t]

Please add link(s) to PRs actually implementing this feature.