Source: vLLM docs: Engine Arguments

Official reference of vLLM engine/server flags and their defaults.

Contents
  • What: official documentation (latest, read 2026-11-14)
  • Author: vLLM project
  • Date: 2026-11-14 (accessed)
  • Original: vLLM docs: Engine Arguments

Summary

The full list of flags accepted by vllm serve and the LLM class, grouped by model, cache, parallel, scheduler, LoRA and compilation configuration.

Key claims

  • --gpu-memory-utilization: fraction of GPU memory the instance may use; default 0.92 (1.0 with an extensible KV cache).
  • --max-model-len defaults from the model config; --dtype auto keeps BF16 for BF16 models and uses FP16 for FP16/FP32 ones; --kv-cache-dtype auto.
  • --enforce-eager (default off) disables CUDA graphs and torch.compile.
  • --tensor-parallel-size, --pipeline-parallel-size, --data-parallel-size default to 1; --enable-expert-parallel (default off) uses EP instead of TP for MoE layers; --decode-context-parallel-size and --prefill-context-parallel-size shard the KV cache / prefill across ranks.
  • --async-scheduling overlaps CPU scheduling with GPU work.
  • --cpu-offload-gb (weights to CPU, default 0) and --kv-offloading-size (CPU KV buffer, off by default).
  • LoRA: --enable-lora, --max-loras (default 1 per batch), --max-lora-rank (default 16).
  • --seed default 0; --trust-remote-code default off; --served-model-name defaults to the model path.

Feeds

vLLM

Suggest changes →