[Doc]: Performance/Optimization Page doesn't mention Pipeline Parallel Size

[strangiato]

📚 The doc issue

In the Page
https://github.com/vllm-project/vllm/blob/main/docs/source/performance/optimization.md

One of the recommended options includes the following:

Increase tensor_parallel_size. This approach shards model weights, so each GPU has more memory available for KV cache.

This document does not mention increasing pipeline_parallel_size which would also result in the model being sharded across more GPUs so their is more memory available for KV cache.

Suggest a potential alternative/fix

Increase tensor_parallel_size or pipeline_parallel_size (if using Multi-Node Multi-GPU). This approach shards model weights, so each GPU has more memory available for KV cache.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.