AI Potluck
Model components / Inference code

vLLM

vLLM (UC Berkeley / vLLM team)

High-throughput LLM serving engine built on PagedAttention, which treats KV cache as virtual memory pages to near-eliminate memory waste during batched inference. Supports continuous batching, tensor parallelism, speculative decoding, and 40+ model architectures. The de facto standard for self-hosted production model serving, used by Anyscale, Databricks, and most open source inference deployments.

v0.22.0, released May 29, 2026 (GitHub releases). Confirmed to exist as of June 2026.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public
governance
vLLM-project(community,PyTorch-ecosystem adjacent)

Adoption

4 high confidence
4.0

~3M+ PyPI downloads/month as of 2026; de facto standard open inference engine with enterprise adoption (Amazon, Roblox, Stripe). 81.9k GitHub stars corroborates but is not the primary basis. Level 4 (1-10M) on download volume; could exceed if cumulative installs counted.

Capability

5 high confidence
5.0

Reference PagedAttention implementation; used in MLPerf Inference submissions (Red Hat OpenShift AI on GPT-J-6b and llama-2-70b). Broad model/hardware coverage (200+ architectures) per repo. Top-tier among open inference engines.

Unchanged since 2026-06-09 (last edited, not re-checked)