| 1 | #!/bin/bash |
| 2 | # This file demonstrates the KV cache event publishing |
| 3 | # We will launch a vllm instances configured to publish KV cache |
| 4 | # events and launch a simple subscriber to log those events. |
| 5 | |
| 6 | set -xe |
| 7 | |
| 8 | echo "🚧🚧 Warning: The usage of KV cache events is experimental and subject to change 🚧🚧" |
| 9 | sleep 1 |
| 10 | |
| 11 | MODEL_NAME=${HF_MODEL_NAME:-meta-llama/Meta-Llama-3.1-8B-Instruct} |
| 12 | |
| 13 | # Trap the SIGINT signal (triggered by Ctrl+C) |
| 14 | trap 'cleanup' INT |
| 15 | |
| 16 | # Cleanup function |
| 17 | cleanup() { |
| 18 | echo "Caught Ctrl+C, cleaning up..." |
| 19 | # Cleanup commands |
| 20 | pgrep python | xargs kill -9 |
| 21 | pkill -f python |
| 22 | echo "Cleanup complete. Exiting." |
| 23 | exit 0 |
| 24 | } |
| 25 | |
| 26 | export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') |
| 27 | |
| 28 | # a function that waits vLLM server to start |
| 29 | wait_for_server() { |
| 30 | local port=$1 |
| 31 | timeout 1200 bash -c " |
| 32 | until curl -s localhost:${port}/v1/completions > /dev/null; do |
| 33 | sleep 1 |
| 34 | done" && return 0 || return 1 |
| 35 | } |
| 36 | |
| 37 | vllm serve "$MODEL_NAME" \ |
| 38 | --port 8100 \ |
| 39 | --max-model-len 100 \ |
| 40 | --enforce-eager \ |
| 41 | --gpu-memory-utilization 0.8 \ |
| 42 | --trust-remote-code \ |
| 43 | --kv-events-config \ |
| 44 | '{"enable_kv_cache_events": true, "publisher": "zmq", "topic": "kv-events"}' & |
| 45 | |
| 46 | wait_for_server 8100 |
| 47 | |
| 48 | SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" |
| 49 | |
| 50 | python3 "$SCRIPT_DIR/kv_events_subscriber.py" & |
| 51 | sleep 1 |
| 52 | |
| 53 | # serve two example requests |
| 54 | output1=$(curl -X POST -s http://localhost:8100/v1/completions \ |
| 55 | -H "Content-Type: application/json" \ |
| 56 | -d '{ |
| 57 | "model": "'"$MODEL_NAME"'", |
| 58 | "prompt": "Explain quantum computing in simple terms a 5-year-old could understand.", |
| 59 | "max_tokens": 80, |
| 60 | "temperature": 0 |
| 61 | }') |
| 62 | |
| 63 | output2=$(curl -X POST -s http://localhost:8100/v1/completions \ |
| 64 | -H "Content-Type: application/json" \ |
| 65 | -d '{ |
| 66 | "model": "'"$MODEL_NAME"'", |
| 67 | "prompt": "Explain quantum computing in simple terms a 50-year-old could understand.", |
| 68 | "max_tokens": 80, |
| 69 | "temperature": 0 |
| 70 | }') |
| 71 | |
| 72 | # Cleanup commands |
| 73 | pkill -9 -u "$USER" -f python |
| 74 | pkill -9 -u "$USER" -f vllm |
| 75 | |
| 76 | sleep 1 |
| 77 | |
| 78 | echo "Cleaned up" |
| 79 | |
| 80 | # Print the outputs of the curl requests |
| 81 | echo "" |
| 82 | echo "Output of first request: $output1" |
| 83 | echo "Output of second request: $output2" |
| 84 | |
| 85 | echo "🎉🎉 Successfully finished 2 test requests! 🎉🎉" |
| 86 | echo "" |