Service can concurrently process multiple requests

[TomasTomecek]

This is a tracking issue for us to figure out for the service to process multiple requests in parallel "so users wouldn't notice" and we don't need to heavily invest into multiple GPUs

[TomasTomecek]

I started researching vllm since it's one of the engines in InstructLab.

In their readme they state (points related to this issue):

• State-of-the-art serving throughput
• Continuous batching of incoming requests
• Quantizations: GPTQ, AWQ, INT4, INT8, and FP8.
• Tensor parallelism and pipeline parallelism support for distributed inference
• OpenAI-compatible API server

There is also a relevant performance benchmark

We can easily test out their OpenAI webserver in a container:
https://docs.vllm.ai/en/latest/serving/deploying_with_docker.html
This will be one of my next steps.

I love they have metrics baked-in so it will be easy for us to measure how many people use our service: https://docs.vllm.ai/en/v0.6.0/serving/metrics.html

One thing to check is the streaming response, there is an open PR over here: vllm-project/vllm#7648

[TomasTomecek]

My writeup from the initial comparison of both vllm and llama-cpp: https://blog.tomecek.net/post/comparing-llamacpp-vllm/