vLLM

vLLM is an inference and serving engine built around PagedAttention, which applies operating-system paging to the KV cache. Conventional serving reserves contiguous memory for each sequence, and because the cache is large and its final size unknown in advance, vLLM's authors measured 60 to 80 percent of that memory lost to fragmentation and over-reservation. PagedAttention splits the cache into fixed-size blocks and maps logical blocks to non-contiguous physical ones through a block table, allocating on demand, which cuts the waste to under 4 percent and leaves far more room for concurrent sequences. Parallel sampling can share blocks between completions.

It has been a PyTorch Foundation hosted project since May 2025, contributed by UC Berkeley's Sky Computing Lab, so governance sits with a neutral foundation rather than a vendor. The licence is a clean, unmodified Apache-2.0. Hardware coverage is broad: NVIDIA CUDA through Blackwell, AMD ROCm, Intel XPU, CPU and Gaudi, Google TPU, AWS Neuron, ARM and Huawei Ascend. Apple Silicon is served by a separate vLLM-Metal project rather than the main distribution, and Windows is not listed as supported.

The documented limitations matter for anyone serving production traffic. vLLM does not guarantee stable log probabilities across runs, and because batching changes with concurrency, the same prompt can produce different output depending on server load; using float32 or fixing a seed reduces this without eliminating it. One server serves one model, so running several means separate instances plus your own routing. Embeddings extracted from generation models underperform purpose-trained embedding models.