docker run -it --shm-size=500g \
  # 环境变量设置
  -e OMNI_INFER_SCRIPTS=$OMNI_INFER_SCRIPTS \
  -e RANKTABLE_SAVE_PATH={{ ranktable_save_path }} \
  -e MODEL_PATH=$MODEL_PATH \
  -e LOG_PATH=$LOG_PATH \
  # 网络和权限设置
  --net=host \
  --privileged=true \
  -u root \
  -w /data \
  # NPU设备挂载
  --device=/dev/davinci_manager \
  --device=/dev/hisi_hdc \
  --device=/dev/devmm_svm \
  --entrypoint=bash \
  # 挂载目录配置
  -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
  -v /usr/local/dcmi:/usr/local/dcmi \
  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
  -v /etc/ascend_install.info:/etc/ascend_install.info \
  -v /usr/local/sbin:/usr/local/sbin \
  -v /etc/hccn.conf:/etc/hccn.conf \
  -v /usr/bin/hccn_tool:/usr/bin/hccn_tool \
  -v /tmp:/tmp \
  -v /data:/data \
  -v $MODEL_PATH:$MODEL_PATH \
  -v $LOG_PATH:$LOG_PATH \
  -v $SCRIPTS_PATH:$SCRIPTS_PATH \
  -v $CODE_PATH:$CODE_PATH \
  -v {{ ranktable_save_path }}:{{ ranktable_save_path }} \
  -v /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime \

# 启动prefill容器
docker run -it --shm-size=500g [各种挂载参数] -d --name $DOCKER_NAME_P $DOCKER_IMAGE_ID

# 启动decode容器  
docker run -it --shm-size=500g [各种挂载参数] -d --name $DOCKER_NAME_D $DOCKER_IMAGE_ID

# 启动proxy容器
docker run -it --shm-size=500g [各种挂载参数] -d --name $DOCKER_NAME_C $DOCKER_IMAGE_ID

# 同步代码
synchronize: src: "{{ ansible_env.CODE_PATH }}/omniinfer" dest: "{{ ansible_env.CODE_PATH }}/"

# 容器内更新代码
docker exec $DOCKER_NAME_P /bin/bash -c ' \
    cd /data/local_code_path/omniinfer/infer_engines && \
    git config --global --add safe.directory /data/local_code_path/omniinfer/infer_engines/vllm && \
    cd vllm && git checkout -f && \
    cd .. && \
    bash bash_install_code.sh && \
    pip uninstall vllm -y && \
    pip uninstall omniinfer -y && \
    cd vllm && \
    SETUPTOOLS_SCM_PRETEND_VERSION=0.9.0 VLLM_TARGET_DEVICE=empty pip install -e . && \
    cd ../../ && \
    pip install -e . && \
    pip uninstall numpy -y && \
    pip install numpy==1.26'

# 生成prefill ranktable
python ${CODE_PATH}/omniinfer/tools/scripts/pd_ranktable_tools.py \
    --mode gen \
    --prefill-server-list "${PREFILL_SERVER_LIST}" \
    --api-server \
    --save-dir ${PREFILL_RANKTABLE_SAVE_PATH}

# 生成decode ranktable
python ${CODE_PATH}/omniinfer/tools/scripts/pd_ranktable_tools.py \
    --mode gen \
    --decode-server-list ${DECODE_SERVER_LIST} \
    --save-dir ${DECODE_RANKTABLE_SAVE_PATH}

# 生成全局ranktable
python ${CODE_PATH}/omniinfer/tools/scripts/pd_ranktable_tools.py \
    --mode merge-all \
    --api-server-list ${api_server_files} \
    --prefill-server-list ${prefill_local_ranktable_merge} \
    --decode-server-list ${decode_local_ranktable_merge} \
    --save-dir ${RANKTABLE_SAVE_PATH}/global

# 杀死Python进程
ps aux | grep "python" | grep -v "grep" | awk '{print $2}' | xargs kill -9

# 启动prefill服务
bash ${CODE_PATH}/omniinfer/tools/scripts/pd_run.sh \
    --global-rank-table-path "${RANKTABLE_SAVE_PATH}/global/global_ranktable_merge.json" \
    --rank-table-path ${LOCAL_RANKTABLE_FLIE} \
    --local-decode-server-ip-list "$SERVER_IP_LIST" \
    --global-decode-server-ip-list "$SERVER_IP_LIST" \
    --prefill-pod-num ${PREFILL_POD_NUM} \
    --gloo-socket-ifname ${SOCKET_IFNAME} \
    --tp-socket-ifname ${SOCKET_IFNAME} \
    --model-path ${MODEL_PATH} \
    --master-ip ${HOST_IP} \
    --role "prefill" \
    --kv-role "kv_producer" \
    --max-model-len ${MODEL_LEN_MAX_PREFILL} \
    --master-port ${MASTER_PORT} \
    --base-api-port ${API_PORT} \
    --tp ${PREFILL_TENSOR_PARALLEL_SIZE} \
    --ascend-rt-visible-devices "${PREFILL_SERVER_LIST}" \
    --kv-rank ${KV_RANK} \
    --kv-engine-id ${KV_RANK} \
    --kv-parallel-size ${KV_PARALLEL_SIZE} \
    --model-extra-cfg-path ${MODEL_EXTRA_CFG_PATH} \
    --gpu-util ${GPU_UTIL} \
    --vllm-enable-mc2 ${VLLM_ENABLE_MC2} \
    --extra-args "${EXTRA_ARGS}" \
    --hccl-buffsize "${HCCL_BUFFSIZE}" \
    --hccl-op-expansion-mode "${HCCL_OP_EXPANSION_MODE}" \
    --log-dir "${LOG_PATH}"

# 启动decode服务
bash ${CODE_PATH}/omniinfer/tools/scripts/pd_run.sh \
	--global-rank-table-path "${RANKTABLE_SAVE_PATH}/global/global_ranktable_merge.json" \
	--rank-table-path ${LOCAL_RANKTABLE_FLIE} \
	--local-decode-server-ip-list "$SERVER_IP_LIST" \
	--global-decode-server-ip-list "$SERVER_IP_LIST" \
	--prefill-pod-num ${PREFILL_POD_NUM} \
	--gloo-socket-ifname ${SOCKET_IFNAME} \
	--tp-socket-ifname ${SOCKET_IFNAME} \
	--num-servers ${NUM_SERVERS} \
	--num-dp ${dp} \
	--server-offset ${config_dict[$HOST]:-0} \
	--model-path ${MODEL_PATH} \
	--master-ip ${HOST_IP} \
	--role "decode" \
	--kv-role "kv_consumer" \
	--max-model-len ${MODEL_LEN_MAX_DECODE} \
	--master-port ${MASTER_PORT} \
	--base-api-port ${API_PORT} \
	--tp ${DECODE_TENSOR_PARALLEL_SIZE} \
	--kv-rank ${PREFILL_POD_NUM} \
	--kv-engine-id ${PREFILL_POD_NUM} \
	--kv-parallel-size ${KV_PARALLEL_SIZE} \
	--model-extra-cfg-path ${MODEL_EXTRA_CFG_PATH} \
	--gpu-util ${GPU_UTIL} \
	--additional-config "$ADDITIONAL_CONFIG" \
	--vllm-enable-mc2 ${VLLM_ENABLE_MC2} \
	--extra-args "${EXTRA_ARGS}" \
	--hccl-buffsize "${HCCL_BUFFSIZE}" \
	--hccl-op-expansion-mode "${HCCL_OP_EXPANSION_MODE}" \
	--log-dir "${LOG_PATH}"

# 杀死nginx进程
ps aux | grep "nginx" | grep -v "grep" | awk '{print $2}' | xargs kill -9

# 下载nginx
wget --no-check-certificate "https://nginx.org/download/nginx-${NGINX_VERSION}.tar.gz"

# 编译nginx
bash build.sh

# 启动代理服务
bash global_proxy.sh \
	--listen-port "$PROXY_NODE_PORT" \
	--prefill-servers-list "$prefill_result" \
	--decode-servers-list "$decode_result" \
	--log-file /dev/shm/trace/logs/nginx_error.log \
	--log-level notice \
	--core-num 4 \
	--start-core-index 16

# 创建日志目录
mkdir -p "{{ ansible_env.LOG_PATH_IN_EXECUTOR }}/{{ inventory_hostname }}"

# 拉取日志
synchronize: mode: pull src: "{{ ansible_env.LOG_PATH }}" dest: "{{ ansible_env.LOG_PATH_IN_EXECUTOR }}/{{ inventory_hostname }}/"

# 在pd_run.sh中使用ranktable配置分布式通信
--global-rank-table-path "${RANKTABLE_SAVE_PATH}/global/global_ranktable_merge.json" \
--rank-table-path ${LOCAL_RANKTABLE_FLIE}

# 在部署脚本中，ranktable信息被转换为服务器列表
prefill_result="{{ PREFILL_API_SERVER_LIST }}"
decode_result="{{ DECODE_API_SERVER_LIST }}"

# 然后传递给Global_Proxy
bash global_proxy.sh \
  --prefill-servers-list "$prefill_result" \
  --decode-servers-list "$decode_result"

vllm serve ${MODEL_PATH} \
    --trust-remote-code \
    --gpu-memory-utilization ${GPU_UTIL} \
    --block_size 128 \
    --tensor-parallel-size ${PREFILL_TENSOR_PARALLEL_SIZE} \
    --data-parallel-size 1 \
    --data-parallel-size-local 1 \
    --data-parallel-address ${HOST_IP} \
    --data-parallel-rpc-port ${MASTER_PORT} \
    --port ${API_PORT} \
    --served-model-name "deepseek" \
    --max-model-len ${MODEL_LEN_MAX_PREFILL} \
    --enable-mtp \
    --speculative_config '{"method": "mtp", "num_speculative_tokens": 1}' \
    --kv-transfer-config '${KV_TRANSFER_CONFIG}' \
    --additional-config '${ADDITIONAL_CONFIG}' \
    ${EXTRA_ARGS}