vllm-sr-sim

vllm-sr-sim is the maintained fleet simulator for this repository. It sizes heterogeneous GPU fleets, evaluates routing strategies, and exposes a service mode that the dashboard can call across containers.

Repository-maintained docs now live in the website:

• https://vllm-sr.ai/docs/fleet-sim/overview
• https://vllm-sr.ai/docs/fleet-sim/getting-started
• https://vllm-sr.ai/docs/fleet-sim/use-cases

Install

cd src/fleet-sim
pip install -e .

Install the service extras when you want to run the simulator API:

pip install -e .[api]

For local development and tests:

pip install -e .[dev]

CLI

vllm-sr-sim --version

vllm-sr-sim optimize \
  --cdf data/azure_cdf.json \
  --lam 200 --slo 500 --b-short 6144 \
  --verify-top 3 --n-sim-req 30000

vllm-sr-sim whatif \
  --cdf data/azure_cdf.json \
  --lam-range 50 100 200 500 1000 \
  --slo 500 --b-short 6144

vllm-sr-sim serve --host 0.0.0.0 --port 8000

vllm-sr serve also starts vllm-sr-sim by default as a sibling container on the shared runtime network so the dashboard can proxy it without rebuilding the router image.

Layout

• fleet_sim/: simulation engine, optimizers, routing, hardware, workload, and service package
• run_sim.py: unified CLI entrypoint used by vllm-sr-sim
• tests/: simulator and service test coverage
• data/: reference workload traces used by the examples and dashboard integration
• examples/: sample scripts and multi-pool input files

Docs

Long-form simulator docs are maintained in the repository website. Keep the package README focused on installation, CLI usage, and source layout.