987f1db4b0
- 实例 A profile 加 --disable-radix-cache（bench 测量纯净） - 新增实例 B profile（kimi3_pro6000_sglang_tp32ep32_instB，.1-.4） - 新增 deploy_dp2.sh 编排脚本（A + B + router --worker-urls） - 新增 README
94 lines
3.6 KiB
Bash
Executable File
94 lines
3.6 KiB
Bash
Executable File
#!/usr/bin/env bash
|
||
# Kimi-K3 DP=2 部署编排脚本（方案 B：两个独立 TP32×EP32 实例 + router 负载均衡）。
|
||
#
|
||
# 架构:
|
||
# 实例 A: 174.1.60.5~8（TP32×EP32，现有 kimi3 profile）
|
||
# 实例 B: 174.1.60.1~4（TP32×EP32，instB profile）
|
||
# router: sglang_router --worker-urls 指向两个实例的 head（.5:30000 与 .1:30000）
|
||
# 对外入口: .5:32000（router），实例间零通信。
|
||
#
|
||
# 用法:
|
||
# bash deploy_dp2.sh start # 起 实例 A → 实例 B → router
|
||
# bash deploy_dp2.sh stop # 停 router → 实例 B → 实例 A
|
||
# bash deploy_dp2.sh status # 查看状态
|
||
# bash deploy_dp2.sh restart
|
||
#
|
||
# 依赖:
|
||
# - 8 节点镜像 lmsysorg/sglang:kimi-k3、patch_k3_sm120.py 已就位
|
||
# - .5 与 .1 上都有 sskj 仓库（deploy 层在各自 master 本地执行）
|
||
# - 各节点 docker 权限、.5 可 ssh .1
|
||
set -Eeuo pipefail
|
||
|
||
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
||
ROOT_DIR="$(cd "${SCRIPT_DIR}/../../.." && pwd)"
|
||
|
||
IMAGE="lmsysorg/sglang:kimi-k3"
|
||
REPO_PATH="/data/yy/sskj"
|
||
MODEL_PATH="/data/hf_models/Kimi-K3"
|
||
|
||
A_HEAD="174.1.60.5"
|
||
B_HEAD="174.1.60.1"
|
||
ROUTER_IP="174.1.60.5"
|
||
ROUTER_CONTAINER="router_dp2"
|
||
ROUTER_PORT="32000"
|
||
|
||
PROFILE_A="pro6000/kimi3_pro6000_sglang_tp32ep32"
|
||
PROFILE_B="pro6000/kimi3_pro6000_sglang_tp32ep32_instB"
|
||
|
||
log() { echo "[$(date +%H:%M:%S)] $*"; }
|
||
|
||
ssh_node() { ssh -o ConnectTimeout=20 -o StrictHostKeyChecking=no "$@"; }
|
||
|
||
run_deploy_on() {
|
||
local host="$1" profile="$2" container="$3" cmd="$4"
|
||
ssh_node "$host" "cd ${REPO_PATH} && PYTHONPATH=src python3 -m sskj.deploy ${cmd} --profile ${profile} --tp 32 --dp 1 --port 30000 --model-path ${MODEL_PATH} --container-name ${container}"
|
||
}
|
||
|
||
start_instance_a() {
|
||
log "starting 实例 A (prefill .5-.8) on ${A_HEAD}"
|
||
run_deploy_on "$A_HEAD" "$PROFILE_A" "kimi3_pro6000_sglang_tp32ep32" start
|
||
}
|
||
|
||
start_instance_b() {
|
||
log "starting 实例 B (.1-.4) on ${B_HEAD}"
|
||
run_deploy_on "$B_HEAD" "$PROFILE_B" "kimi3_pro6000_sglang_tp32ep32_instB" start
|
||
}
|
||
|
||
start_router() {
|
||
log "starting router (--worker-urls) on ${ROUTER_IP}:${ROUTER_PORT}"
|
||
ssh_node "$ROUTER_IP" "docker rm -f ${ROUTER_CONTAINER} >/dev/null 2>&1 || true; docker run -d --name ${ROUTER_CONTAINER} --network host ${IMAGE} python3 -m sglang_router.launch_router --worker-urls http://${B_HEAD}:30000 http://${A_HEAD}:30000 --host 0.0.0.0 --port ${ROUTER_PORT}" >/dev/null
|
||
sleep 5
|
||
if ssh_node "$ROUTER_IP" "ss -tlnp 2>/dev/null | grep -q :${ROUTER_PORT}"; then
|
||
log "router is up (${ROUTER_PORT})"
|
||
else
|
||
log "WARN: router ${ROUTER_PORT} not listening"
|
||
fi
|
||
}
|
||
|
||
stop_all() {
|
||
log "stopping router"
|
||
ssh_node "$ROUTER_IP" "docker rm -f ${ROUTER_CONTAINER} >/dev/null 2>&1 || true" || true
|
||
log "stopping 实例 B"
|
||
run_deploy_on "$B_HEAD" "$PROFILE_B" "kimi3_pro6000_sglang_tp32ep32_instB" stop || true
|
||
log "stopping 实例 A"
|
||
run_deploy_on "$A_HEAD" "$PROFILE_A" "kimi3_pro6000_sglang_tp32ep32" stop || true
|
||
log "all stopped"
|
||
}
|
||
|
||
status_all() {
|
||
echo "=== 实例 A (.5-.8) ==="
|
||
run_deploy_on "$A_HEAD" "$PROFILE_A" "kimi3_pro6000_sglang_tp32ep32" status || true
|
||
echo "=== 实例 B (.1-.4) ==="
|
||
run_deploy_on "$B_HEAD" "$PROFILE_B" "kimi3_pro6000_sglang_tp32ep32_instB" status || true
|
||
echo "=== router ==="
|
||
ssh_node "$ROUTER_IP" "docker ps --format '{{.Names}} {{.Status}}' | grep ${ROUTER_CONTAINER} || echo stopped"
|
||
}
|
||
|
||
case "${1:-}" in
|
||
start) start_instance_a; sleep 20; start_instance_b; sleep 20; start_router ;;
|
||
stop) stop_all ;;
|
||
status) status_all ;;
|
||
restart) stop_all; sleep 5; start_instance_a; sleep 20; start_instance_b; sleep 20; start_router ;;
|
||
*) echo "用法: $0 {start|stop|status|restart}"; exit 1 ;;
|
||
esac
|
||
echo "done." |