Common vLLM flags (quick reference)

Full reference: vLLM engine arguments.

Flag	Meaning
--tensor-parallel-size	Split the model across N GPUs (TP)
--max-model-len	Max context length to support (bounds KV-cache size)
--gpu-memory-utilization	Fraction of GPU memory vLLM may use
--max-num-seqs	Max concurrent requests
--max-num-batched-tokens	Prefill chunk size / per-step token budget
--enable-prefix-caching	Reuse KV cache for shared prompt prefixes
--kv-cache-dtype fp8	Store KV cache in 8-bit
--quantization	Weight quantization scheme to use
--speculative-config	Enable/configure speculative decoding
--enforce-eager	Disable CUDA graphs (slower, more compatible)
--trust-remote-code	Allow running model-supplied Python code (needed by some architectures)
--served-model-name	The model name clients use in API calls
--hf-overrides	Patch fields in the model’s HF config at load time
--override-generation-config	Patch the model’s default sampling params (temperature, top_p, …)
--tool-call-parser / --reasoning-parser	Which parser converts the model’s raw markup into API fields
--attention-backend	Which attention kernel library to use
--language-model-only	Skip loading the vision part of a multimodal model