| 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_PORT="${PREFILL_PORT:-19535}" |
| 15 | DECODE_PORT="${DECODE_PORT:-19536}" |
| 16 | PROXY_PORT="${PROXY_PORT:-10001}" |
| 17 | |
| 18 | GPU_E="${GPU_E:-2}" |
| 19 | GPU_P="${GPU_P:-2}" |
| 20 | GPU_D="${GPU_D:-3}" |
| 21 | |
| 22 | # Device platform and affinity env name. |
| 23 | # DEVICE_PLATFORM supports: cuda, xpu |
| 24 | DEVICE_PLATFORM="${DEVICE_PLATFORM:-cuda}" |
| 25 | if [[ -z "${DEVICE_AFFINITY_ENV:-}" ]]; then |
| 26 | if [[ "${DEVICE_PLATFORM,,}" == "xpu" ]]; then |
| 27 | DEVICE_AFFINITY_ENV="ZE_AFFINITY_MASK" |
| 28 | else |
| 29 | DEVICE_AFFINITY_ENV="CUDA_VISIBLE_DEVICES" |
| 30 | fi |
| 31 | fi |
| 32 | |
| 33 | EC_SHARED_STORAGE_PATH="${EC_SHARED_STORAGE_PATH:-/tmp/ec_cache}" |
| 34 | TIMEOUT_SECONDS="${TIMEOUT_SECONDS:-12000}" # wait_for_server timeout |
| 35 | |
| 36 | NUM_PROMPTS="${NUM_PROMPTS:-100}" # number of prompts to send in benchmark |
| 37 | |
| 38 | # Serve args |
| 39 | GPU_MEMORY_UTILIZATION_E="${GPU_MEMORY_UTILIZATION_E:-0.01}" |
| 40 | GPU_MEMORY_UTILIZATION_P="${GPU_MEMORY_UTILIZATION_P:-0.7}" |
| 41 | GPU_MEMORY_UTILIZATION_D="${GPU_MEMORY_UTILIZATION_D:-0.7}" |
| 42 | MAX_NUM_SEQS="${MAX_NUM_SEQS:-128}" |
| 43 | MAX_MODEL_LEN="${MAX_MODEL_LEN:-32768}" |
| 44 | |
| 45 | export UCX_TLS=all |
| 46 | export UCX_NET_DEVICES=all |
| 47 | |
| 48 | ############################################################################### |
| 49 | # Helpers |
| 50 | ############################################################################### |
| 51 | # Find the git repository root directory |
| 52 | GIT_ROOT=$(git rev-parse --show-toplevel) |
| 53 | |
| 54 | START_TIME=$(date +"%Y%m%d_%H%M%S") |
| 55 | ENC_LOG=$LOG_PATH/encoder_${START_TIME}.log |
| 56 | P_LOG=$LOG_PATH/p_${START_TIME}.log |
| 57 | D_LOG=$LOG_PATH/d_${START_TIME}.log |
| 58 | PROXY_LOG=$LOG_PATH/proxy_${START_TIME}.log |
| 59 | |
| 60 | wait_for_server() { |
| 61 | local port=$1 |
| 62 | timeout "$TIMEOUT_SECONDS" bash -c " |
| 63 | until curl -s localhost:$port/v1/chat/completions > /dev/null; do |
| 64 | sleep 1 |
| 65 | done" && return 0 || return 1 |
| 66 | } |
| 67 | |
| 68 | # Cleanup function |
| 69 | cleanup() { |
| 70 | echo "Stopping everything…" |
| 71 | trap - INT TERM USR1 # prevent re-entrancy |
| 72 | |
| 73 | # Kill all tracked PIDs |
| 74 | for pid in "${PIDS[@]}"; do |
| 75 | if kill -0 "$pid" 2>/dev/null; then |
| 76 | echo "Killing process $pid" |
| 77 | kill "$pid" 2>/dev/null |
| 78 | fi |
| 79 | done |
| 80 | |
| 81 | # Wait a moment for graceful shutdown |
| 82 | sleep 2 |
| 83 | |
| 84 | # Force kill any remaining processes |
| 85 | for pid in "${PIDS[@]}"; do |
| 86 | if kill -0 "$pid" 2>/dev/null; then |
| 87 | echo "Force killing process $pid" |
| 88 | kill -9 "$pid" 2>/dev/null |
| 89 | fi |
| 90 | done |
| 91 | |
| 92 | # Kill the entire process group as backup |
| 93 | kill -- -$$ 2>/dev/null |
| 94 | |
| 95 | echo "All processes stopped." |
| 96 | exit 0 |
| 97 | } |
| 98 | |
| 99 | trap cleanup INT |
| 100 | trap cleanup USR1 |
| 101 | trap cleanup TERM |
| 102 | |
| 103 | # clear previous cache |
| 104 | echo "remove previous ec cache folder" |
| 105 | rm -rf "$EC_SHARED_STORAGE_PATH" |
| 106 | |
| 107 | echo "make ec cache folder" |
| 108 | mkdir -p "$EC_SHARED_STORAGE_PATH" |
| 109 | |
| 110 | ############################################################################### |
| 111 | # Encoder worker |
| 112 | ############################################################################### |
| 113 | env "$DEVICE_AFFINITY_ENV=$GPU_E" vllm serve "$MODEL" \ |
| 114 | --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION_E" \ |
| 115 | --port "$ENCODE_PORT" \ |
| 116 | --enforce-eager \ |
| 117 | --enable-request-id-headers \ |
| 118 | --no-enable-prefix-caching \ |
| 119 | --max-num-batched-tokens 114688 \ |
| 120 | --max-num-seqs "$MAX_NUM_SEQS" \ |
| 121 | --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \ |
| 122 | --ec-transfer-config '{ |
| 123 | "ec_connector": "ECExampleConnector", |
| 124 | "ec_role": "ec_producer", |
| 125 | "ec_connector_extra_config": { |
| 126 | "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'" |
| 127 | } |
| 128 | }' \ |
| 129 | >"${ENC_LOG}" 2>&1 & |
| 130 | |
| 131 | PIDS+=($!) |
| 132 | |
| 133 | ############################################################################### |
| 134 | # Prefill worker |
| 135 | ############################################################################### |
| 136 | env "$DEVICE_AFFINITY_ENV=$GPU_P" \ |
| 137 | UCX_NET_DEVICES=all \ |
| 138 | VLLM_NIXL_SIDE_CHANNEL_PORT=5559 \ |
| 139 | vllm serve "$MODEL" \ |
| 140 | --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION_P" \ |
| 141 | --port "$PREFILL_PORT" \ |
| 142 | --enforce-eager \ |
| 143 | --enable-request-id-headers \ |
| 144 | --max-num-seqs "$MAX_NUM_SEQS" \ |
| 145 | --max-model-len "$MAX_MODEL_LEN" \ |
| 146 | --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \ |
| 147 | --ec-transfer-config '{ |
| 148 | "ec_connector": "ECExampleConnector", |
| 149 | "ec_role": "ec_consumer", |
| 150 | "ec_connector_extra_config": { |
| 151 | "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'" |
| 152 | } |
| 153 | }' \ |
| 154 | --kv-transfer-config '{ |
| 155 | "kv_connector": "NixlConnector", |
| 156 | "kv_role": "kv_producer" |
| 157 | }' \ |
| 158 | >"${P_LOG}" 2>&1 & |
| 159 | |
| 160 | PIDS+=($!) |
| 161 | |
| 162 | ############################################################################### |
| 163 | # Decode worker |
| 164 | ############################################################################### |
| 165 | env "$DEVICE_AFFINITY_ENV=$GPU_D" \ |
| 166 | UCX_NET_DEVICES=all \ |
| 167 | VLLM_NIXL_SIDE_CHANNEL_PORT=6000 \ |
| 168 | vllm serve "$MODEL" \ |
| 169 | --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION_D" \ |
| 170 | --port "$DECODE_PORT" \ |
| 171 | --enforce-eager \ |
| 172 | --enable-request-id-headers \ |
| 173 | --max-num-seqs "$MAX_NUM_SEQS" \ |
| 174 | --max-model-len "$MAX_MODEL_LEN" \ |
| 175 | --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \ |
| 176 | --kv-transfer-config '{ |
| 177 | "kv_connector": "NixlConnector", |
| 178 | "kv_role": "kv_consumer" |
| 179 | }' \ |
| 180 | >"${D_LOG}" 2>&1 & |
| 181 | |
| 182 | PIDS+=($!) |
| 183 | |
| 184 | # Wait for workers |
| 185 | wait_for_server "$ENCODE_PORT" |
| 186 | wait_for_server "$PREFILL_PORT" |
| 187 | wait_for_server "$DECODE_PORT" |
| 188 | |
| 189 | ############################################################################### |
| 190 | # Proxy |
| 191 | ############################################################################### |
| 192 | PROXY_ARGS=() |
| 193 | if [[ "${DYNAMIC_REGISTRATION:-0}" == "1" ]]; then |
| 194 | : "${ADMIN_API_KEY:?Set ADMIN_API_KEY for dynamic registration}" |
| 195 | export ADMIN_API_KEY |
| 196 | PROXY_ARGS=(--dynamic-registration) |
| 197 | else |
| 198 | PROXY_ARGS=( |
| 199 | --encode-servers-urls "http://localhost:$ENCODE_PORT" |
| 200 | --prefill-servers-urls "http://localhost:$PREFILL_PORT" |
| 201 | --decode-servers-urls "http://localhost:$DECODE_PORT" |
| 202 | ) |
| 203 | fi |
| 204 | python disagg_epd_proxy.py \ |
| 205 | --host "0.0.0.0" \ |
| 206 | --port "$PROXY_PORT" \ |
| 207 | "${PROXY_ARGS[@]}" \ |
| 208 | >"${PROXY_LOG}" 2>&1 & |
| 209 | |
| 210 | PIDS+=($!) |
| 211 | |
| 212 | wait_for_server "$PROXY_PORT" |
| 213 | if [[ "${DYNAMIC_REGISTRATION:-0}" == "1" ]]; then |
| 214 | for endpoint in "encode:$ENCODE_PORT" "prefill:$PREFILL_PORT" "decode:$DECODE_PORT"; do |
| 215 | curl --fail-with-body "http://127.0.0.1:$PROXY_PORT/instances" \ |
| 216 | -H "X-API-Key: $ADMIN_API_KEY" -H 'Content-Type: application/json' \ |
| 217 | -d "{\"role\":\"${endpoint%%:*}\",\"url\":\"http://127.0.0.1:${endpoint#*:}\"}" |
| 218 | done |
| 219 | fi |
| 220 | echo "All services are up!" |
| 221 | |
| 222 | ############################################################################### |
| 223 | # Benchmark |
| 224 | ############################################################################### |
| 225 | echo "Running benchmark (stream)..." |
| 226 | vllm bench serve \ |
| 227 | --model "$MODEL" \ |
| 228 | --backend openai-chat \ |
| 229 | --endpoint /v1/chat/completions \ |
| 230 | --dataset-name hf \ |
| 231 | --dataset-path lmarena-ai/VisionArena-Chat \ |
| 232 | --seed 0 \ |
| 233 | --num-prompts "$NUM_PROMPTS" \ |
| 234 | --port "$PROXY_PORT" |
| 235 | |
| 236 | PIDS+=($!) |
| 237 | |
| 238 | ############################################################################### |
| 239 | # Single request with local image |
| 240 | ############################################################################### |
| 241 | echo "Running single request with local image (non-stream)..." |
| 242 | curl http://127.0.0.1:"${PROXY_PORT}"/v1/chat/completions \ |
| 243 | -H "Content-Type: application/json" \ |
| 244 | -d '{ |
| 245 | "model": "'"${MODEL}"'", |
| 246 | "messages": [ |
| 247 | {"role": "system", "content": "You are a helpful assistant."}, |
| 248 | {"role": "user", "content": [ |
| 249 | {"type": "image_url", "image_url": {"url": "file://'"${GIT_ROOT}"'/tests/v1/ec_connector/integration/hato.jpg"}}, |
| 250 | {"type": "text", "text": "What is in this image?"} |
| 251 | ]} |
| 252 | ] |
| 253 | }' |
| 254 | |
| 255 | |
| 256 | # cleanup |
| 257 | echo "cleanup..." |
| 258 | cleanup |