[V1] [7/N] API Server: Multiprocessing Detokenizer [ DO NOT MERGE ]

[robertgshaw2-redhat]

NOTES:

• Adding a 3rd process hurts performance
• Current main API server is within 3% of the LLM performance already
• Performance Benchmarks

So, we will abandon the 3 process architecure and go forward with the 2 process architecture

SUMMARY:

• Splitting up [V1] Improve AsyncLLM and API Server #11237
• Update shutdown logic to use weakref._finalizer
• Split detokenizer and API server into separate processes

PERFORMANCE:

Server

• Launch:

vllm serve $MODEL --no-enable-prefix-caching --gpu-memory-utilization 0.98 --max-num-batched-tokens 8192 --disable-log-requests

• Benchmark:

python3 benchmark_serving.py --model $MODEL --dataset-name sonnet --dataset-path sonnet.txt --sonnet-input-len 250 --sonnet-output-len 200

• Result:
• main: 55.99 requests/sec
• pr: 54.05 requests/sec

AsyncLLM

python3 benchmarks/benchmark_throughput.py --model $MODEL --input-len 230 --output-len 200 --no-enable-prefix-caching --gpu-memory-utilization 0.98 --max-num-batched-tokens 8192 --async-engine

>> Throughput: 55.10 requests/s

• main

python3 benchmarks/benchmark_throughput.py --model $MODEL --input-len 230 --output-len 200 --no-enable-prefix-caching --gpu-memory-utilization 0.98 --max-num-batched-tokens 8192 --async-engine

>> Throughput: 55.26 requests/s

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[robertgshaw2-redhat]

Closing as the implementation did not lead to performance gains