| 1 | #!/bin/bash |
| 2 | set -euo pipefail |
| 3 | |
| 4 | declare -a PIDS=() |
| 5 | |
| 6 | ############################################################################### |
| 7 | # Configuration -- override via env before running |
| 8 | ############################################################################### |
| 9 | MODEL="${MODEL:-Qwen/Qwen2.5-VL-3B-Instruct}" |
| 10 | LOG_PATH="${LOG_PATH:-./logs}" |
| 11 | mkdir -p "$LOG_PATH" |
| 12 | |
| 13 | ENCODE_PORT="${ENCODE_PORT:-19534}" |
| 14 | PREFILL_DECODE_PORT="${PREFILL_DECODE_PORT:-19535}" |
| 15 | PROXY_PORT="${PROXY_PORT:-10001}" |
| 16 | |
| 17 | GPU_E="${GPU_E:-0}" |
| 18 | GPU_PD="${GPU_PD:-1}" |
| 19 | |
| 20 | # Device platform and affinity env name. |
| 21 | # DEVICE_PLATFORM supports: cuda, xpu |
| 22 | DEVICE_PLATFORM="${DEVICE_PLATFORM:-cuda}" |
| 23 | if [[ -z "${DEVICE_AFFINITY_ENV:-}" ]]; then |
| 24 | if [[ "${DEVICE_PLATFORM,,}" == "xpu" ]]; then |
| 25 | DEVICE_AFFINITY_ENV="ZE_AFFINITY_MASK" |
| 26 | else |
| 27 | DEVICE_AFFINITY_ENV="CUDA_VISIBLE_DEVICES" |
| 28 | fi |
| 29 | fi |
| 30 | |
| 31 | EC_SHARED_STORAGE_PATH="${EC_SHARED_STORAGE_PATH:-/tmp/ec_cache}" |
| 32 | TIMEOUT_SECONDS="${TIMEOUT_SECONDS:-300}" # wait_for_server timeout |
| 33 | |
| 34 | NUM_PROMPTS="${NUM_PROMPTS:-100}" # number of prompts to send in benchmark |
| 35 | |
| 36 | # Serve args |
| 37 | GPU_MEMORY_UTILIZATION_E="${GPU_MEMORY_UTILIZATION_E:-0.01}" |
| 38 | GPU_MEMORY_UTILIZATION_PD="${GPU_MEMORY_UTILIZATION_PD:-0.7}" |
| 39 | MAX_NUM_SEQS="${MAX_NUM_SEQS:-128}" |
| 40 | MAX_MODEL_LEN="${MAX_MODEL_LEN:-32768}" |
| 41 | |
| 42 | ############################################################################### |
| 43 | # Helpers |
| 44 | ############################################################################### |
| 45 | # Find the git repository root directory |
| 46 | GIT_ROOT=$(git rev-parse --show-toplevel) |
| 47 | |
| 48 | START_TIME=$(date +"%Y%m%d_%H%M%S") |
| 49 | ENC_LOG=$LOG_PATH/encoder_${START_TIME}.log |
| 50 | PD_LOG=$LOG_PATH/pd_${START_TIME}.log |
| 51 | PROXY_LOG=$LOG_PATH/proxy_${START_TIME}.log |
| 52 | |
| 53 | wait_for_server() { |
| 54 | local port=$1 |
| 55 | timeout "$TIMEOUT_SECONDS" bash -c " |
| 56 | until curl -s localhost:$port/v1/chat/completions > /dev/null; do |
| 57 | sleep 1 |
| 58 | done" && return 0 || return 1 |
| 59 | } |
| 60 | |
| 61 | # Cleanup function |
| 62 | cleanup() { |
| 63 | echo "Stopping everything…" |
| 64 | trap - INT TERM USR1 # prevent re-entrancy |
| 65 | |
| 66 | # Kill all tracked PIDs |
| 67 | for pid in "${PIDS[@]}"; do |
| 68 | if kill -0 "$pid" 2>/dev/null; then |
| 69 | echo "Killing process $pid" |
| 70 | kill "$pid" 2>/dev/null |
| 71 | fi |
| 72 | done |
| 73 | |
| 74 | # Wait a moment for graceful shutdown |
| 75 | sleep 2 |
| 76 | |
| 77 | # Force kill any remaining processes |
| 78 | for pid in "${PIDS[@]}"; do |
| 79 | if kill -0 "$pid" 2>/dev/null; then |
| 80 | echo "Force killing process $pid" |
| 81 | kill -9 "$pid" 2>/dev/null |
| 82 | fi |
| 83 | done |
| 84 | |
| 85 | # Kill the entire process group as backup |
| 86 | kill -- -$$ 2>/dev/null |
| 87 | |
| 88 | echo "All processes stopped." |
| 89 | exit 0 |
| 90 | } |
| 91 | |
| 92 | trap cleanup INT |
| 93 | trap cleanup USR1 |
| 94 | trap cleanup TERM |
| 95 | |
| 96 | # clear previous cache |
| 97 | echo "remove previous ec cache folder" |
| 98 | rm -rf "$EC_SHARED_STORAGE_PATH" |
| 99 | |
| 100 | echo "make ec cache folder" |
| 101 | mkdir -p "$EC_SHARED_STORAGE_PATH" |
| 102 | |
| 103 | ############################################################################### |
| 104 | # Encoder worker |
| 105 | ############################################################################### |
| 106 | env "$DEVICE_AFFINITY_ENV=$GPU_E" vllm serve "$MODEL" \ |
| 107 | --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION_E" \ |
| 108 | --port "$ENCODE_PORT" \ |
| 109 | --enforce-eager \ |
| 110 | --enable-request-id-headers \ |
| 111 | --no-enable-prefix-caching \ |
| 112 | --max-num-batched-tokens 114688 \ |
| 113 | --max-num-seqs "$MAX_NUM_SEQS" \ |
| 114 | --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \ |
| 115 | --ec-transfer-config '{ |
| 116 | "ec_connector": "ECExampleConnector", |
| 117 | "ec_role": "ec_producer", |
| 118 | "ec_connector_extra_config": { |
| 119 | "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'" |
| 120 | } |
| 121 | }' \ |
| 122 | >"${ENC_LOG}" 2>&1 & |
| 123 | |
| 124 | PIDS+=($!) |
| 125 | |
| 126 | ############################################################################### |
| 127 | # Prefill+Decode worker |
| 128 | ############################################################################### |
| 129 | env "$DEVICE_AFFINITY_ENV=$GPU_PD" vllm serve "$MODEL" \ |
| 130 | --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION_PD" \ |
| 131 | --port "$PREFILL_DECODE_PORT" \ |
| 132 | --enforce-eager \ |
| 133 | --enable-request-id-headers \ |
| 134 | --max-num-seqs "$MAX_NUM_SEQS" \ |
| 135 | --max-model-len "$MAX_MODEL_LEN" \ |
| 136 | --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \ |
| 137 | --ec-transfer-config '{ |
| 138 | "ec_connector": "ECExampleConnector", |
| 139 | "ec_role": "ec_consumer", |
| 140 | "ec_connector_extra_config": { |
| 141 | "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'" |
| 142 | } |
| 143 | }' \ |
| 144 | >"${PD_LOG}" 2>&1 & |
| 145 | |
| 146 | PIDS+=($!) |
| 147 | |
| 148 | # Wait for workers |
| 149 | wait_for_server "$ENCODE_PORT" |
| 150 | wait_for_server "$PREFILL_DECODE_PORT" |
| 151 | |
| 152 | ############################################################################### |
| 153 | # Proxy |
| 154 | ############################################################################### |
| 155 | PROXY_ARGS=() |
| 156 | if [[ "${DYNAMIC_REGISTRATION:-0}" == "1" ]]; then |
| 157 | : "${ADMIN_API_KEY:?Set ADMIN_API_KEY for dynamic registration}" |
| 158 | export ADMIN_API_KEY |
| 159 | PROXY_ARGS=(--dynamic-registration) |
| 160 | else |
| 161 | PROXY_ARGS=( |
| 162 | --encode-servers-urls "http://localhost:$ENCODE_PORT" |
| 163 | --decode-servers-urls "http://localhost:$PREFILL_DECODE_PORT" |
| 164 | ) |
| 165 | fi |
| 166 | python disagg_epd_proxy.py \ |
| 167 | --host "0.0.0.0" \ |
| 168 | --port "$PROXY_PORT" \ |
| 169 | "${PROXY_ARGS[@]}" \ |
| 170 | >"${PROXY_LOG}" 2>&1 & |
| 171 | |
| 172 | PIDS+=($!) |
| 173 | |
| 174 | wait_for_server "$PROXY_PORT" |
| 175 | if [[ "${DYNAMIC_REGISTRATION:-0}" == "1" ]]; then |
| 176 | for endpoint in "encode:$ENCODE_PORT" "prefill_decode:$PREFILL_DECODE_PORT"; do |
| 177 | curl --fail-with-body "http://127.0.0.1:$PROXY_PORT/instances" \ |
| 178 | -H "X-API-Key: $ADMIN_API_KEY" -H 'Content-Type: application/json' \ |
| 179 | -d "{\"role\":\"${endpoint%%:*}\",\"url\":\"http://127.0.0.1:${endpoint#*:}\"}" |
| 180 | done |
| 181 | fi |
| 182 | echo "All services are up!" |
| 183 | |
| 184 | ############################################################################### |
| 185 | # Benchmark |
| 186 | ############################################################################### |
| 187 | echo "Running benchmark (stream)..." |
| 188 | vllm bench serve \ |
| 189 | --model "$MODEL" \ |
| 190 | --backend openai-chat \ |
| 191 | --endpoint /v1/chat/completions \ |
| 192 | --dataset-name hf \ |
| 193 | --dataset-path lmarena-ai/VisionArena-Chat \ |
| 194 | --seed 0 \ |
| 195 | --num-prompts "$NUM_PROMPTS" \ |
| 196 | --port "$PROXY_PORT" |
| 197 | |
| 198 | PIDS+=($!) |
| 199 | |
| 200 | ############################################################################### |
| 201 | # Single request with local image |
| 202 | ############################################################################### |
| 203 | echo "Running single request with local image (non-stream)..." |
| 204 | curl http://127.0.0.1:"${PROXY_PORT}"/v1/chat/completions \ |
| 205 | -H "Content-Type: application/json" \ |
| 206 | -d '{ |
| 207 | "model": "'"${MODEL}"'", |
| 208 | "messages": [ |
| 209 | {"role": "system", "content": "You are a helpful assistant."}, |
| 210 | {"role": "user", "content": [ |
| 211 | {"type": "image_url", "image_url": {"url": "file://'"${GIT_ROOT}"'/tests/v1/ec_connector/integration/hato.jpg"}}, |
| 212 | {"type": "text", "text": "What is in this image?"} |
| 213 | ]} |
| 214 | ] |
| 215 | }' |
| 216 | |
| 217 | |
| 218 | # cleanup |
| 219 | echo "cleanup..." |
| 220 | cleanup |