vLLM Environment Variables

LISA Serve supports configuring vLLM model serving through environment variables. These variables allow you to control performance, memory usage, parallelization, and advanced features when deploying models with vLLM.

• NOTE: Standard vLLM environment variables are supported and passed directly into the VLLM container. See vLLM's documentation

Core Performance & Memory

Variable	Description	Default	Example
VLLM_GPU_MEMORY_UTILIZATION	Fraction of GPU memory to use (0.0-1.0)	0.9	0.85
VLLM_MAX_MODEL_LEN	Maximum context length override	Auto	4096
MAX_TOTAL_TOKENS	Legacy alias for VLLM_MAX_MODEL_LEN	Auto	4096

Model Format & Loading

Variable	Description	Default	Example
VLLM_DTYPE	Model precision	auto	half, float16, bfloat16, float32
VLLM_QUANTIZATION	Quantization method	-	awq, gptq, squeezellm, fp8
VLLM_TRUST_REMOTE_CODE	Allow custom model code execution	false	true

Performance Tuning

Variable	Description	Default	Example
VLLM_MAX_NUM_BATCHED_TOKENS	Maximum tokens per batch	Auto	8192
VLLM_MAX_NUM_SEQS	Maximum concurrent sequences	256	128, 512
VLLM_ENABLE_PREFIX_CACHING	Enable prefix caching for repeated prompts	false	true
VLLM_ENABLE_CHUNKED_PREFILL	Enable chunked prefill	false	true

Parallel Processing

Variable	Description	Default	Example
VLLM_TENSOR_PARALLEL_SIZE	Split model across N GPUs	1	2, 4, 8

Tool Calling / Function Calling

Variable	Description	Default	Example
VLLM_ENABLE_AUTO_TOOL_CHOICE	Enable automatic tool choice routing	false	true
VLLM_TOOL_CALL_PARSER	Tool call parser implementation	-	hermes, mistral, llama3_json, qwen

Note: Tool calling requires both VLLM_ENABLE_AUTO_TOOL_CHOICE=true and specifying an appropriate VLLM_TOOL_CALL_PARSER for your model. See vLLM Tool Calling Documentation for details.

Reference

For more details on vLLM configuration, see the official vLLM documentation.