[XPU] Add test script of PD disaggregation (#36434)

Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>

tests/v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh

+#!/bin/bash
+set -e
+
+# Hosts / ports
+PREFILL_HOST=${PREFILL_HOST:-"localhost"}
+PREFILL_PORT=${PREFILL_PORT:-8100}
+PREFILL_NIXL_SIDE_PORT=${PREFILL_NIXL_SIDE_PORT:-5577}
+DECODE_HOST=${DECODE_HOST:-"localhost"}
+DECODE_PORT=${DECODE_PORT:-8200}
+PROXY_HOST=${PROXY_HOST:-"localhost"}
+PROXY_PORT=${PROXY_PORT:-8192}
+BASELINE_HOST=${BASELINE_HOST:-"localhost"}
+BASELINE_PORT=${BASELINE_PORT:-9290}
+
+# Model to run.
+MODEL_NAME=${MODEL_NAME:-"Qwen/Qwen3-0.6B"}
+MAX_MODEL_LEN=${MAX_MODEL_LEN:-1024}
+BLOCK_SIZE=${BLOCK_SIZE:-64}
+PREFILLER_TP_SIZE=${PREFILLER_TP_SIZE:-1}
+DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
+KV_BUFFER_DEVICE=${KV_BUFFER_DEVICE:-"xpu"}
+GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.8}
+
+generate_affinity_mask() {
+  local count=$1
+  local start=${2:-0}
+  local mask=""
+  local i
+
+  for ((i=0; i<count; i++)); do
+    local device=$((start + i))
+    if [[ -z "${mask}" ]]; then
+      mask="${device}"
+    else
+      mask="${mask},${device}"
+    fi
+  done
+
+  echo "${mask}"
+}
+
+PREFILLER_ZE_AFFINITY_MASK=${PREFILLER_ZE_AFFINITY_MASK:-$(generate_affinity_mask "${PREFILLER_TP_SIZE}" 0)}
+DECODER_ZE_AFFINITY_MASK=${DECODER_ZE_AFFINITY_MASK:-$(generate_affinity_mask "${DECODER_TP_SIZE}" "${PREFILLER_TP_SIZE}")}
+
+
+# execution env
+GIT_ROOT=$(git rev-parse --show-toplevel)
+EXP_ROOT="${GIT_ROOT}/tests/v1/kv_connector/nixl_integration"
+
+OUTPUT_FILE=${OUTPUT_FILE:-"${EXP_ROOT}/.xpu_accuracy_test_outputs.txt"}
+
+# Trap the SIGINT signal (triggered by Ctrl+C)
+trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
+
+cleanup() {
+  echo "Cleaning up any running vLLM instances..."
+  pkill -f "vllm serve" || true
+  sleep 2
+}
+
+wait_for_server() {
+  local host=$1
+  local port=$2
+  timeout 1200 bash -c "
+    until curl -s ${host}:${port}/v1/completions > /dev/null; do
+      sleep 1
+    done" && return 0 || return 1
+}
+
+launch_baseline() {
+  BASELINE_BASE_CMD="
+  ZE_AFFINITY_MASK=0 \
+  VLLM_WORKER_MULTIPROC_METHOD=spawn \
+  VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
+      --host ${BASELINE_HOST} \
+      --port ${BASELINE_PORT} \
+      --max-model-len ${MAX_MODEL_LEN}\
+      --seed 42 \
+      -tp 1 \
+      --block-size ${BLOCK_SIZE} \
+      --gpu-memory-utilization ${GPU_MEMORY_UTILIZATION} \
+      --dtype float16 \
+      --enforce-eager"
+  echo ${BASELINE_BASE_CMD}      
+  bash -c "${BASELINE_BASE_CMD}" &
+  sleep 10
+  wait_for_server ${BASELINE_HOST} ${BASELINE_PORT}
+}
+
+launch_pd() {
+  PREFILL_BASE_CMD="
+  ZE_AFFINITY_MASK=${PREFILLER_ZE_AFFINITY_MASK} \
+  VLLM_MULTIPROC_EXECUTE_MODEL_TIMEOUT_S=200 \
+  VLLM_NIXL_SIDE_CHANNEL_HOST=${PREFILL_HOST} \
+  VLLM_NIXL_SIDE_CHANNEL_PORT=${PREFILL_NIXL_SIDE_PORT} \
+  VLLM_WORKER_MULTIPROC_METHOD=spawn \
+  VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
+      --host ${PREFILL_HOST} \
+      --port ${PREFILL_PORT} \
+      --max-model-len ${MAX_MODEL_LEN}\
+      --seed 42 \
+      --block-size ${BLOCK_SIZE} \
+      --enforce-eager \
+      --dtype float16 \
+      -tp ${PREFILLER_TP_SIZE} \
+      --gpu-memory-utilization ${GPU_MEMORY_UTILIZATION} \
+      --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\",\"kv_buffer_device\":\"$KV_BUFFER_DEVICE\"}'"
+
+
+  DECODE_BASE_CMD="
+  ZE_AFFINITY_MASK=${DECODER_ZE_AFFINITY_MASK} \
+  VLLM_MULTIPROC_EXECUTE_MODEL_TIMEOUT_S=200 \
+  VLLM_WORKER_MULTIPROC_METHOD=spawn \
+  VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
+      --host ${DECODE_HOST} \
+      --port ${DECODE_PORT} \
+      --max-model-len ${MAX_MODEL_LEN}\
+      --seed 42 \
+      --block-size ${BLOCK_SIZE} \
+      --enforce-eager \
+      -tp ${DECODER_TP_SIZE} \
+      --dtype float16 \
+      --gpu-memory-utilization ${GPU_MEMORY_UTILIZATION} \
+      --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\",\"kv_buffer_device\":\"$KV_BUFFER_DEVICE\"}'"
+
+  echo ${PREFILL_BASE_CMD}
+  echo ${DECODE_BASE_CMD}
+  sleep 2
+
+  # execute on hosts
+  bash -c "${PREFILL_BASE_CMD}" &
+  bash -c "${DECODE_BASE_CMD}" &
+  sleep 1
+  wait_for_server ${PREFILL_HOST} ${PREFILL_PORT}
+  sleep 1
+  wait_for_server ${DECODE_HOST} ${DECODE_PORT}
+  sleep 1
+}
+
+launch_pd_proxy(){
+  PROXY_BASE_CMD="
+  python3 ${EXP_ROOT}/toy_proxy_server.py \
+  --prefiller-host ${PREFILL_HOST} --prefiller-port ${PREFILL_PORT} \
+  --decoder-host ${DECODE_HOST} --decoder-port ${DECODE_PORT} \
+  --host=${PROXY_HOST} --port ${PROXY_PORT}"
+  echo ${PROXY_BASE_CMD} 
+  bash -c "${PROXY_BASE_CMD}" &
+  sleep 2
+}
+
+run_tests(){
+  local service_url=$1
+  local mode=$2
+  python3 ${EXP_ROOT}/test_disagg_accuracy.py --service_url=${service_url} --model_name=${MODEL_NAME} --mode=${mode} --file_name=${OUTPUT_FILE}
+}
+
+
+# run non-disagg. baseline & save outputs
+launch_baseline
+run_tests "http://${BASELINE_HOST}:${BASELINE_PORT}" "baseline"
+cleanup
+sleep 10
+
+
+# run disagg. & do exact-match with the outputs from baseline
+launch_pd
+launch_pd_proxy
+run_tests "http://${PROXY_HOST}:${PROXY_PORT}" "disagg"
+echo "-----P/D success----"
+
+rm ${OUTPUT_FILE}
+cleanup
+
+exit 0