Quick start
You need Linux or WSL, an NVIDIA GPU, and a local model directory. Native Windows can inspect configurations and reports but cannot run vLLM.
1. Prepare the tools
Create and activate an isolated Python 3.11 or 3.12 environment:
2. Install the complete runtime
This installs vTune, vLLM, and GuideLLM together. Use plain pip install vtune
only on machines that will inspect configurations, artifacts, and reports
without launching experiments. See installation choices.
3. Create experiment.yaml
experiment:
name: first-run
server:
model: /models/opt-125m
gpu-memory-utilization: 0.8
tune:
max-num-seqs:
values: [8, 16]
enforce-eager:
values: [true, false]
max-num-batched-tokens:
min: 4096
max: 8192
step: 4096
env:
CUDA_VISIBLE_DEVICES: "0"
tune_env:
VLLM_USE_FLASHINFER_SAMPLER:
values: ["0", "1"]
benchmark:
runs:
- name: throughput
profile:
kind: throughput
max_concurrency: 16
constraints:
- kind: max_requests
count: 10
data:
- kind: synthetic_text
prompt_tokens: 32
output_tokens: 16
optimization:
maximize: output_tokens_per_second
sampler: tpe
trials: 2
4. Run
Open report.html in the printed run directory when the run completes. Add
--verbose to stream server and benchmark logs while trials execute.
values accepts categorical strings, numbers, or booleans. Numeric parameters
can instead use inclusive min, max, and step ranges. The same two forms
work under tune_env; selected environment values are converted to strings.
Next, learn how configuration maps to vLLM and GuideLLM.