# vLLM Engine Arguments vLLM engine arguments, flags, options for serving models on vLLM.
ArgumentExample and use-case
--gpu-memory-utilizationDefault 0.9. How much VRAM usage vLLM can use. Reduce if going out of memory. Try setting this to 0.95 or 0.97.
--max-model-lenSet maximum sequence length. Reduce this if going out of memory! For example set --max-model-len 32768 to use only 32K sequence lengths.
--quantizationUse fp8 for dynamic float8 quantization. Use this in tandem with --kv-cache-dtype fp8 to enable float8 KV cache as well.
--kv-cache-dtypeUse fp8 for float8 KV cache to reduce memory usage by 50%.
--portDefault is 8000. How to access vLLM's localhost ie http://localhost:8000
--api-keyOptional - Set the password (or no password) to access the model.
--tensor-parallel-sizeDefault is 1. Splits model across tensors. Set this to how many GPUs you are using - if you have 4, set this to 4. 8, then 8. You should have NCCL, otherwise this might be slow.
--pipeline-parallel-sizeDefault is 1. Splits model across layers. Use this with --pipeline-parallel-size where TP is used within each node, and PP is used across multi-node setups (set PP to number of nodes)
--enable-loraEnables LoRA serving. Useful for serving Unsloth finetuned LoRAs.
--max-lorasHow many LoRAs you want to serve at 1 time. Set this to 1 for 1 LoRA, or say 16. This is a queue so LoRAs can be hot-swapped.
--max-lora-rankMaximum rank of all LoRAs. Possible choices are 8, 16, 32, 64, 128, 256, 320, 512
--dtypeAllows auto, bfloat16, float16 Float8 and other quantizations use a different flag - see --quantization
--tokenizerSpecify the tokenizer path like unsloth/gpt-oss-20b if the served model has a different tokenizer.
--hf-tokenAdd your HuggingFace token if needed for gated models
--swap-spaceDefault is 4GB. CPU offloading usage. Reduce if you have VRAM, or increase for low memory GPUs.
--seedDefault is 0 for vLLM
--disable-log-statsDisables logging like throughput, server requests.
--enforce-eagerDisables compilation. Faster to load, but slower for inference.
--disable-cascade-attnUseful for Reinforcement Learning runs for vLLM < 0.11.0, as Cascade Attention was slightly buggy on A100 GPUs (Unsloth fixes this)
### :tada:Float8 Quantization For example to host Llama 3.3 70B Instruct (supports 128K context length) with Float8 KV Cache and quantization, try: ```bash vllm serve unsloth/Llama-3.3-70B-Instruct \ --quantization fp8 \ --kv-cache-dtype fp8 --gpu-memory-utilization 0.97 \ --max-model-len 65536 ``` ### :shaved\_ice:LoRA Hot Swapping / Dynamic LoRAs To enable LoRA serving for at most 4 LoRAs at 1 time (these are hot swapped / changed), first set the environment flag to allow hot swapping: See our [lora-hot-swapping-guide](https://unsloth.ai/docs/basics/inference-and-deployment/vllm-guide/lora-hot-swapping-guide "mention") for more details. --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://unsloth.ai/docs/basics/inference-and-deployment/vllm-guide/vllm-engine-arguments.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.