[Perf] API-server scaleout with many-to-many server-engine comms

[njhill]

This is a follow-on from #15977.

A new --api-server-count arg to vllm serve can be used to specify an arbitrary number of API servers to run. When used in conjunction with --data-parallel-size there's all-to-all zmq-based communication between API servers and data parallel engines.

It works with multi-node as described in #15977. All of the API servers run on the head node.

A separate "coordinator" process is now used for DP>1. This is responsible for ensuring that the engines run in tandem, and for publishing real-time request count information (and later likely other engine state info) back to the api server(s) for load balancing purposes.

More design discussion: https://docs.google.com/document/d/10jhCNxJYvsUhtMtiMAaW2MxU5LU8HVje2pGDnj49gH4/edit?tab=t.0

Performance now scales much better with DP size. Observe TTFT in particular below.

Benchmark with 2xA100, llama-3.2-1B, share-gpt with request rate 120 req/sec:

DP=2 before

============ Serving Benchmark Result ============
Successful requests:                     10000     
Benchmark duration (s):                  130.74    
Total input tokens:                      2206428   
Total generated tokens:                  1994815   
Request throughput (req/s):              76.49     
Output token throughput (tok/s):         15258.46  
Total Token throughput (tok/s):          32135.56  
---------------Time to First Token----------------
Mean TTFT (ms):                          13176.40  
Median TTFT (ms):                        13953.03  
P99 TTFT (ms):                           26842.02  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          19.83     
Median TPOT (ms):                        22.28     
P99 TPOT (ms):                           36.10     
---------------Inter-token Latency----------------
Mean ITL (ms):                           24.19     
Median ITL (ms):                         21.98     
P99 ITL (ms):                            81.11     
==================================================

DP=2 with --api-server-count=2

============ Serving Benchmark Result ============
Successful requests:                     10000     
Benchmark duration (s):                  116.84    
Total input tokens:                      2206428   
Total generated tokens:                  1994815   
Request throughput (req/s):              85.59     
Output token throughput (tok/s):         17073.43  
Total Token throughput (tok/s):          35958.03  
---------------Time to First Token----------------
Mean TTFT (ms):                          67.54     
Median TTFT (ms):                        60.81     
P99 TTFT (ms):                           329.10    
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          23.90     
Median TPOT (ms):                        24.13     
P99 TPOT (ms):                           36.89     
---------------Inter-token Latency----------------
Mean ITL (ms):                           23.62     
Median ITL (ms):                         22.14     
P99 ITL (ms):                            51.75     
==================================================

This is working functionally but there are still a number of tasks remaining:

• Initial benchmark results for DP=1 and multiple API servers are disappointing - I am looking into why this currently hurts ITL and throughput slightly (though TTFT slightly improves).
• (medium) Multiple api servers don't currently work properly with metrics publishing/logging. I have discussed this with @markmc but it needs a bit more work. @kouroshHakha is helping to look at this, I will add some more notes below.
• (small) The multi-modal embeddings cache currently won't work with DP and/or mutli-API so will need to be auto-disabled when dp > 1 and/or api-server-count > 1. Hopefully the scale-out should hide the performance downsides to that however (discussed this with @ywang96 and @DarkLight1337).
• (small) When there are many API servers, a lot of the startup logs are duplicated. We probably want to suppress some of these.
• (tbd) Need to look into implications for LoRA adapter loading.
• (medium) Some more work on error handling and clean shutdown with the new process topologies.
• (medium) Full test coverage of the various permutations.

Follow-on work (not for this PR):

• Rework how the multi-modal feature cache is implemented to make it compatible with the any-to-any process architecture.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[yinghai]

How does run_rpc work if we want to bcast this to each engine and run it exactly once? How to guarantee that each engine core runs it in lock step if we want?

[yinghai]

There isn't a lot of work in apiserver that needs multiprocessing right? It's mostly async_llm, most specifically MM data handling that needs scale out?

[njhill]

I think all the CI issues are fixed and remaining failures should be unrelated, we should let it finish though.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @njhill.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[lgeiger]

I am now seeing the following warnings on main when running some test e.g.:

pytest tests/v1/engine/test_async_llm.py::test_load -s

ERROR 05-31 01:04:14 [prometheus.py:77] Error during metrics cleanup: expected str, bytes or os.PathLike object, not NoneType

This warning is thrown during the prometheus cleanup though not sure where this is coming from exactly.

[lgeiger]

Looks like the error comes from this prometheus function

[njhill]

Thanks @lgeiger. I think the message is harmless but I'll fix this.

[njhill]

@lgeiger fixed in #18992.