[Usage]: Is pipeline parallelism supported on machines that are not in the same local network?

[oldcpple]

How would you like to use vllm

Hi there, since the communications between nodes are done by NCCL(which typically relies on RDMA I guess), I wonder if I can setup an inference pipeline with machines from different networks, for example, one on Google Cloud and another on AWS Cloud, through vLLM's pipeline parallelism？
Thanks a lot if anyone can answer this.

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.

[noooop]

Do you really want to do this?

A typical vllm step takes about 20ms, and copying an intermediate result (a large tenser) over the network is very slow.

And now vllm is scheduled synchronously, so the delay in network transmission of intermediate results will greatly reduce GPU utilization, increase latency, and reduce throughput.

You can pay attention to progress of Disaggregated prefilling

It seems to be asynchronous, that Awesome

[lihuahua123]

Do you really want to do this?

A typical vllm step takes about 20ms, and copying an intermediate result (a large tenser) over the network is very slow.

And now vllm is scheduled synchronously, so the delay in network transmission of intermediate results will greatly reduce GPU utilization, increase latency, and reduce throughput.

You can pay attention to progress of Disaggregated prefilling

It seems to be asynchronous, that Awesome

Is the Disaggregated prefilling better when the network between the machines is slow? And it also need double memory i guess?

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!