By the numbers

A quantitative snapshot of the vLLM repository.

Data collected on 2026-04-30 from the main branch at commit efb4cdf2b.

Size

xychart-beta horizontal
    title "Lines by area (rough order of magnitude)"
    x-axis ["vllm/ Python", "csrc/ C++/CUDA", "tests/", "docs/", "examples/", "benchmarks/"]
    y-axis "Lines" 0 --> 700000
    bar [614883, 82337, 250000, 30000, 25000, 25000]

Area	Files	Notes
vllm/ Python	1,637	The primary package. ~615 K lines of Python.
vllm/model_executor/models/	293	One Python file per model architecture (some have multiple variants).
csrc/ C++/CUDA	229	~82 K lines of C++/CUDA/CUTLASS
tests/	1,169	pytest files
docs/	211	Markdown pages for https://docs.vllm.ai
requirements/	~30	Pinned and unpinned dependency lists

Top-level directory	Purpose
vllm/	Python package source
csrc/	C++/CUDA kernels
tests/	pytest suites
docs/	mkdocs-material site
examples/	Runnable examples
benchmarks/	Throughput / latency / serving harnesses
tools/	Build & dev scripts
cmake/	CMake helpers used by setup.py
requirements/	Per-platform dep lists
docker/	Dockerfiles
.buildkite/	The primary CI pipeline (Buildkite)
.github/	GitHub workflows for pre-commit, issue templates, codeowners

The largest single Python files give a feel for where complexity concentrates:

File	Lines
vllm/v1/worker/gpu_model_runner.py	7,070
vllm/engine/arg_utils.py	2,476
vllm/v1/core/sched/scheduler.py	2,308
vllm/v1/engine/core.py	2,145
vllm/entrypoints/llm.py	1,934
vllm/config/vllm.py	~2,200
vllm/config/model.py	~2,200
vllm/v1/engine/async_llm.py	1,065
vllm/entrypoints/openai/api_server.py	719

Activity

Window	Commits
All time (since Feb 2023)	16,223
Calendar 2023	613
Calendar 2024	3,352
Since 2025-01-01	12,246
Last 90 days	2,780

vLLM's development pace roughly quadrupled in early 2024 and has continued to accelerate. Around 30 % of the project's lifetime commits landed in the most recent 12 months.

Bot-attributed commits

vLLM does receive automated contributions, but they are a small slice of the total:

Bot signal	Count
Author = dependabot[bot]	21
Commit body contains Co-authored-by: (any)	5,858
Commit body contains Co-authored-by: ...[bot]	~50

The 5,858 figure is dominated by human collaborators co-authoring each other's commits — bot collaboration is rare and is mostly dependabot[bot], gemini-code-assist[bot], github-advanced-security[bot] (Copilot Autofix), codeflash-ai[bot], mergify[bot], and the occasional copilot-swe-agent[bot]. This count is a lower bound on AI-assisted work since inline AI tools (Copilot, Cursor, etc.) leave no trace in git history. AGENTS.md explicitly forbids "pure code-agent PRs" — every change must have a human submitter.

Complexity

Hot files (largest by line count) by area:

Area	File	Lines
Forward pass orchestration	vllm/v1/worker/gpu_model_runner.py	7,070
Argument parsing	vllm/engine/arg_utils.py	2,476
Scheduler	vllm/v1/core/sched/scheduler.py	2,308
EngineCore loop	vllm/v1/engine/core.py	2,145
Offline LLM API	vllm/entrypoints/llm.py	1,934
Config aggregator	vllm/config/vllm.py	~2,200
Model config	vllm/config/model.py	~2,200
Compilation config	vllm/config/compilation.py	~1,400
Parallel state	vllm/distributed/parallel_state.py	~1,800
FlashInfer attention backend	vllm/v1/attention/backends/flashinfer.py	~1,800
Fused MoE GPU kernel wrapper	vllm/model_executor/layers/fused_moe/fused_moe.py	~2,000
Linear layer (TP/PP/quant variants)	vllm/model_executor/layers/linear.py	~1,500
Modular kernel framework	vllm/model_executor/layers/fused_moe/modular_kernel.py	~1,500
Custom ops surface	vllm/_custom_ops.py	~3,000
AITER ops surface (ROCm)	vllm/_aiter_ops.py	~2,000
Environment overrides	vllm/env_override.py, vllm/envs.py	~3,200

The largest C++/CUDA TUs are in csrc/quantization/ (CUTLASS-based GEMMs), csrc/moe/ (block-scaled MoE kernels), and csrc/attention/ (custom attention variants).

Per-area ownership signals

Recent contributor activity by area (from git log on main, last ~6 months):

Area	Recent contributors (top, anonymized to first names where ambiguous)
Engine (V1)	Nick Hill, Robert Shaw, Cyrus Leung, Woosuk Kwon
Scheduler / KV cache	Cyrus Leung, Cody Yu, Andreas Karatzas
Attention backends	Lucas Wilkinson, Matthew Bonanni, Wentao Ye
Model implementations	Isotr0py, Roger Wang, Cyrus Leung, Harry Mellor
Quantization	Michael Goin, rasmith, Lucas Wilkinson
ROCm / AMD	Micah Williamson, rasmith, Charlie Fu
MoE / EPLB	Wentao Ye, Wensheng Tang, Tyler Michael Smith
Tooling / CI	Harry Mellor, Kevin H. Luu
Frontend (OpenAI server)	Harry Mellor, Cyrus Leung, Nick Hill

For per-page bylines, see the maintainers page.