Ascend-SACT/DeepSeek-V3.1-Terminus-w8a8c8
模型介绍文件和版本Pull Requests讨论分析
下载使用量0

基于MindIE最新版本（2.2.RC1），使用昇腾A2双机16卡部署DeepSeek-V3.1-Terminus-w8a8c8模型。本方案包含双机配置文件，并针对不同上下文长度场景提供性能最优配置。

1. 模型概述及场景

2025年9月22日，DeepSeek-V3.1升级至DeepSeek-V3.1-Terminus版本。此次更新在保留模型既有能力的前提下，针对用户反馈的问题进行了优化，具体包括：

① 语言一致性：减轻了中英文混杂、偶发异常字符等现象；

② Agent能力：进一步提升了Code Agent与Search Agent的性能表现。

DeepSeek-V3.1-Terminus的输出效果较上一版本更为稳定，尤其适用于长文本处理任务与代码生成场景。

以下是该模型在昇腾NPU卡上的部署指南。

模型权重下载地址：https://modelscope.cn/models/Eco-Tech/DeepSeek-V3.1-Terminus-w8a8c8-mtp-QuaRot

2. 准备运行环境

2.1 硬件环境准备

设备数量
Atlas 800I/800T A2(910B 64G*8)2
交换机（CE9855/9860/XH9210）1
400G-2*200G光纤4

服务器需具备网络权限，以便下载镜像和权重文件。 若无法联网，请提前下载相关文件并上传至服务器。

2.2 软件环境准备

软件版本
Python3.11.10
PyTorch7.2.0
MindIE2.2.RC1
Ascend HDK25.2.1
CANN8.3.RC2

3. 运行指导

3.1 准备工作

请提前将所有昇腾服务器上的HDK和CANN升级至最新版本，升级驱动参考。

注意：选择arm版本，先升级固件，再升级驱动。

升级过程中会导致业务中断，请提前做好变更规划。

3.2 下载推理镜像和模型权重

3.2.1 下载推理镜像

通过ssh工具分别登录至昇腾服务器后台，执行下载操作。 登录昇腾社区/昇腾镜像仓库，下载镜像前需申请权限， 待权限申请通过后，根据指南下载对应镜像文件。 权限申请通过后，可使用docker pull命令下载MindIE推理框架镜像：

docker pull --platform=arm64 swr.cn-south-1.myhuaweicloud.com/ascendhub/mindie:2.2.RC1-800I-A2-py311-openeuler24.03-lts

3.2.2 下载模型权重

安装modelscope，通过网络下载模型权重。

pip install modelscope

通过modelscope将模型权重文件下载至服务器目录，下载前需确保两台服务器均有足够空间存储权重文件，文件总大小为647GB。

下载过程耗时较长，具体时间取决于网络状况，建议采用nohup方式进行下载。

nohup modelscope download --model Eco-Tech/DeepSeek-V3.1-Terminus-w8a8c8 --local_dir /ai > ds3.1t-w8a8c8-download.log 2>&1 &

下载后修改模型目录下的config.json，将第54行的"torch_dtype"的值"bfloat16"改为"float16"。

在一台服务器上下载好之后通过scp复制到另一台服务器相同目录下，以节省时间。

scp /ai/DeepSeek-V3.1-Terminus-w8a8c8 {user}@{另一台服务器ip}:/ai

3.3 创建推理容器

使用docker run命令创建容器并挂载8张npu卡。

docker run -itd --privileged=true --name=ds3.1t-w8a8c8 --net=host --shm-size 800g \
--device=/dev/davinci0 \
--device=/dev/davinci1 \
--device=/dev/davinci2 \
--device=/dev/davinci3 \
--device=/dev/davinci4 \
--device=/dev/davinci5 \
--device=/dev/davinci6 \
--device=/dev/davinci7 \
--device=/dev/davinci_manager \
--device=/dev/hisi_hdc \
--device /dev/devmm_svm \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \
-v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \
-v /usr/local/sbin:/usr/local/sbin \
-v /etc/hccn.conf:/etc/hccn.conf \
-v /home/:/home/ \
-v /ai:/ai \
swr.cn-south-1.myhuaweicloud.com/ascendhub/mindie:2.2.RC1-800I-A2-py311-openeuler24.03-lts \
/bin/bash

注意在本次部署模型所用到的2台昇腾服务器上都执行该操作。上面命令中挂载的/ai目录中存放的是模型的权重文件，视具体情况修改。

3.4 启动推理服务

在2台昇腾服务器上执行下列操作： 进入容器：

docker exec -it ds3.1t-w8a8c8 bash

3.4.1 创建ranktable.json配置文件

该文件指定了每个节点的每个卡的id和卡通信的ip以及序号：

vi ranktable.json

输入以下内容:

{
	"version": "1.0",
	"server_count": "2",
	"server_list": [
		{
			"server_id": "{节点1 IP}",
			"container_ip": "{节点1 IP}",
			"device": [
				{
					"device_id": "0",
					"device_ip": "{卡通信ip}",
					"rank_id": "0"
				},
				{
					"device_id": "1",
					"device_ip": "{卡通信ip}",
					"rank_id": "1"
				},
				{
					"device_id": "2",
					"device_ip": "{卡通信ip}",
					"rank_id": "2"
				},
				{
					"device_id": "3",
					"device_ip": "{卡通信ip}",
					"rank_id": "3"
				},
				{
					"device_id": "4",
					"device_ip": "{卡通信ip}",
					"rank_id": "4"
				},
				{
					"device_id": "5",
					"device_ip": "{卡通信ip}",
					"rank_id": "5"
				},
				{
					"device_id": "6",
					"device_ip": "{卡通信ip}",
					"rank_id": "6"
				},
				{
					"device_id": "7",
					"device_ip": "{卡通信ip}",
					"rank_id": "7"
				}
			],
			"host_nic_ip": "reserve"
		},
		{
			"server_id": "{节点2 IP}",
			"container_ip": "{节点2 IP}",
			"device": [
				{
					"device_id": "0",
					"device_ip": "{卡通信ip}",
					"rank_id": "8"
				},
				{
					"device_id": "1",
					"device_ip": "{卡通信ip}",
					"rank_id": "9"
				},
				{
					"device_id": "2",
					"device_ip": "{卡通信ip}",
					"rank_id": "10"
				},
				{
					"device_id": "3",
					"device_ip": "{卡通信ip}",
					"rank_id": "11"
				},
				{
					"device_id": "4",
					"device_ip": "{卡通信ip}",
					"rank_id": "12"
				},
				{
					"device_id": "5",
					"device_ip": "{卡通信ip}",
					"rank_id": "13"
				},
				{
					"device_id": "6",
					"device_ip": "{卡通信ip}",
					"rank_id": "14"
				},
				{
					"device_id": "7",
					"device_ip": "{卡通信ip}",
					"rank_id": "15"
				}
			],
			"host_nic_ip": "reserve"
		}
	],
	"status": "completed"
}

配置文件中的device_ip通过在每个节点上执行下面的命令查看：

for i in {0..7};do hccn_tool -i $i -ip -g; done

ranktable.json配置文件内容在2个节点上一致。记住该文件全路径，后面配置时要用。

3.4.2 创建环境变量文件

vi env.sh

输入以下内容

source /usr/local/Ascend/mindie/set_env.sh
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/atb-models/set_env.sh
export MIES_CONTAINER_IP={本机IP}
export RANK_TABLE_FILE={ranktable.json文件全路径}
export MASTER_IP={主节点IP}
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ATB_WORKSPACE_MEM_ALLOC_ALG_TYPE=3
export ATB_WORKSPACE_MEM_ALLOC_GLOBAL=1
export HCCL_OP_EXPANSION_MODE="AIV"
export NPU_MEMORY_FRACTION=0.96
export ATB_LLM_HCCL_ENABLE=1
export INF_NAN_MODE_ENABLE=1
# 8244特性防止oom开关
export ATB_LAYER_INTERNAL_TENSOR_REUSE=1
export HCCL_CONNECT_TIMEOUT=3600
export WORLD_SIZE=16
export HCCL_EXEC_TIMEOUT=0
# A3到15，A3双机到33
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export ATB_OPERATION_EXECUTE_ASYNC=1
export ATB_LLM_ENABLE_AUTO_TRANSPOSE=0
export HCCL_RDMA_PCIE_DIRECT_POST_NOSTRICT=TRUE
export HCCL_INTRA_PCIE_ENABLE=1
export HCCL_INTRA_ROCE_ENABLE=0
export HCCL_BUFFSIZE=64
# 异步双发
export MINDIE_ASYNC_SCHEDULING_ENABLE=1
# jemalloc优化，社区镜像中的路径为/usr/lib64/libjemalloc.so.2；研发镜像中的路径为/usr/lib/aarch64-linux-gnu/libjemalloc.so.2
#export LD_PRELOAD={libjemalloc.so在本环境上的路径}
# 队列优化特性
export TASK_QUEUE_ENABLE=1

for var in $(compgen -e | grep 'STDOUT$'); do
  export "$var=0"
done
 
for var in $(compgen -e | grep 'LOG_TO_FILE$'); do
  export "$var=0"
done
 
# 遇到config.json文件权限问题
find /usr/local/lib/python3.11/site-packages/mindie*  -name  config.json |xargs chmod -R 640
#chmod -R 640 {双机ranktable的路径}
 
# 日志开关
export MINDIE_LOG_TO_STDOUT=0
export MINDIE_LOG_TO_FILE=1
export MINDIE_LOG_LEVEL=info
 
export OMP_NUM_THREADS=10

#HCCL显存分配策略，配置后不会跑着跑着继续申请显存。
export HCCL_ALGO="level0:NA;level1:pipeline"

主节点IP一般取第一个节点的IP。保存文件后，使用下面的命令加载环境变量。

source env.sh

等待环境变量加载成功。

3.4.3 修改配置config.json文件

cd /usr/local/Ascend/mindie/latest/mindie-service/conf
vi config.json

如果是短上下文（<32K）推理，使用下面的配置：

{
    "Version" : "1.0.0",

    "ServerConfig" :
    {
        "ipAddress" : "{主节点IP}",
        "managementIpAddress" : "{主节点IP}",
        "port" : 1025,
        "managementPort" : 1026,
        "metricsPort" : 1027,
        "allowAllZeroIpListening" : false,
        "maxLinkNum" : 1000,
        "httpsEnabled" : false,
        "fullTextEnabled" : false,
        "tlsCaPath" : "security/ca/",
        "tlsCaFile" : ["ca.pem"],
        "tlsCert" : "security/certs/server.pem",
        "tlsPk" : "security/keys/server.key.pem",
        "tlsPkPwd" : "security/pass/key_pwd.txt",
        "tlsCrlPath" : "security/certs/",
        "tlsCrlFiles" : ["server_crl.pem"],
        "managementTlsCaFile" : ["management_ca.pem"],
        "managementTlsCert" : "security/certs/management/server.pem",
        "managementTlsPk" : "security/keys/management/server.key.pem",
        "managementTlsPkPwd" : "security/pass/management/key_pwd.txt",
        "managementTlsCrlPath" : "security/management/certs/",
        "managementTlsCrlFiles" : ["server_crl.pem"],
        "kmcKsfMaster" : "tools/pmt/master/ksfa",
        "kmcKsfStandby" : "tools/pmt/standby/ksfb",
        "inferMode" : "standard",
        "interCommTLSEnabled" : false,
        "interCommPort" : 1121,
        "interCommTlsCaPath" : "security/grpc/ca/",
        "interCommTlsCaFiles" : ["ca.pem"],
        "interCommTlsCert" : "security/grpc/certs/server.pem",
        "interCommPk" : "security/grpc/keys/server.key.pem",
        "interCommPkPwd" : "security/grpc/pass/key_pwd.txt",
        "interCommTlsCrlPath" : "security/grpc/certs/",
        "interCommTlsCrlFiles" : ["server_crl.pem"],
        "openAiSupport" : "vllm",
        "tokenTimeout" : 3600,
        "e2eTimeout" : 65535,
        "distDPServerEnabled":false
    },

    "BackendConfig" : {
        "backendName" : "mindieservice_llm_engine",
        "modelInstanceNumber" : 1,
        "npuDeviceIds" : [[0,1,2,3,4,5,6,7]],
        "tokenizerProcessNumber" : 8,
        "multiNodesInferEnabled" : true,
        "multiNodesInferPort" : 1120,
        "interNodeTLSEnabled" : false,
        "interNodeTlsCaPath" : "security/grpc/ca/",
        "interNodeTlsCaFiles" : ["ca.pem"],
        "interNodeTlsCert" : "security/grpc/certs/server.pem",
        "interNodeTlsPk" : "security/grpc/keys/server.key.pem",
        "interNodeTlsPkPwd" : "security/grpc/pass/mindie_server_key_pwd.txt",
        "interNodeTlsCrlPath" : "security/grpc/certs/",
        "interNodeTlsCrlFiles" : ["server_crl.pem"],
        "interNodeKmcKsfMaster" : "tools/pmt/master/ksfa",
        "interNodeKmcKsfStandby" : "tools/pmt/standby/ksfb",
        "kvPoolConfig" : {"backend":"", "configPath":""},
        "ModelDeployConfig" :
        {
            "maxSeqLen" : 32768,
            "maxInputTokenLen" : 32768,
            "truncation" : false,
            "ModelConfig" : [
                {
                    "modelInstanceType" : "Standard",
                    "modelName" : "ds3.1t-w8a8c8",
                    "modelWeightPath" : "/ai/DeepSeek-V3.1-Terminus-w8a8c8/",
                    "worldSize" : 8,
                    "cpuMemSize" : 5,
                    "npuMemSize" : -1,
                    "backendType" : "atb",
                    "trustRemoteCode" : false,
                    "async_scheduler_wait_time": 120,
                    "kv_trans_timeout": 10,
                    "kv_link_timeout": 1080,
                    "dp": 2,
                    "sp": 1,
                    "tp": 8,
                    "cp": 1,
                    "moe_ep": 4,
                    "moe_tp": 4,
                    "plugin_params":"{\"plugin_type\":\"mtp\",\"num_speculative_tokens\": 1}",
                    "models": {
                        "deepseekv2": {
                            "enable_mlapo_prefetch": true,
                            "kv_cache_options": {"enable_nz": true}
                       }
                   },
                   "llm": {
                        "parallel_options": {
                        "dense_mlp_local_tp": 16
                        }
                }
                }
            ]
        },

        "ScheduleConfig" :
        {
            "templateType" : "Standard",
            "templateName" : "Standard_LLM",
            "cacheBlockSize" : 128,

            "maxPrefillBatchSize" : 50,
            "maxPrefillTokens" : 32768,
            "prefillTimeMsPerReq" : 150,
            "prefillPolicyType" : 0,

            "decodeTimeMsPerReq" : 50,
            "decodePolicyType" : 0,

            "maxBatchSize" : 200,
            "maxIterTimes" : 32768,
            "maxPreemptCount" : 0,
            "supportSelectBatch" : false,
            "maxQueueDelayMicroseconds" : 5000
        }
    },

    "LogConfig": {
        "dynamicLogLevel" : "",
        "dynamicLogLevelValidHours" : 2,
        "dynamicLogLevelValidTime" : ""
    }
}

如果要支持到128K上下文推理，使用下面的配置：

{
    "Version" : "1.0.0",

    "ServerConfig" :
    {
        "ipAddress" : "{主节点IP}",
        "managementIpAddress" : "{主节点IP}",
        "port" : 1025,
        "managementPort" : 1026,
        "metricsPort" : 1027,
        "allowAllZeroIpListening" : false,
        "maxLinkNum" : 1000,
        "httpsEnabled" : false,
        "fullTextEnabled" : false,
        "tlsCaPath" : "security/ca/",
        "tlsCaFile" : ["ca.pem"],
        "tlsCert" : "security/certs/server.pem",
        "tlsPk" : "security/keys/server.key.pem",
        "tlsPkPwd" : "security/pass/key_pwd.txt",
        "tlsCrlPath" : "security/certs/",
        "tlsCrlFiles" : ["server_crl.pem"],
        "managementTlsCaFile" : ["management_ca.pem"],
        "managementTlsCert" : "security/certs/management/server.pem",
        "managementTlsPk" : "security/keys/management/server.key.pem",
        "managementTlsPkPwd" : "security/pass/management/key_pwd.txt",
        "managementTlsCrlPath" : "security/management/certs/",
        "managementTlsCrlFiles" : ["server_crl.pem"],
        "kmcKsfMaster" : "tools/pmt/master/ksfa",
        "kmcKsfStandby" : "tools/pmt/standby/ksfb",
        "inferMode" : "standard",
        "interCommTLSEnabled" : false,
        "interCommPort" : 1121,
        "interCommTlsCaPath" : "security/grpc/ca/",
        "interCommTlsCaFiles" : ["ca.pem"],
        "interCommTlsCert" : "security/grpc/certs/server.pem",
        "interCommPk" : "security/grpc/keys/server.key.pem",
        "interCommPkPwd" : "security/grpc/pass/key_pwd.txt",
        "interCommTlsCrlPath" : "security/grpc/certs/",
        "interCommTlsCrlFiles" : ["server_crl.pem"],
        "openAiSupport" : "vllm",
        "tokenTimeout" : 3600,
        "e2eTimeout" : 65535,
        "distDPServerEnabled":false
    },

    "BackendConfig" : {
        "backendName" : "mindieservice_llm_engine",
        "modelInstanceNumber" : 1,
        "npuDeviceIds" : [[0,1,2,3,4,5,6,7]],
        "tokenizerProcessNumber" : 8,
        "multiNodesInferEnabled" : true,
        "multiNodesInferPort" : 1120,
        "interNodeTLSEnabled" : false,
        "interNodeTlsCaPath" : "security/grpc/ca/",
        "interNodeTlsCaFiles" : ["ca.pem"],
        "interNodeTlsCert" : "security/grpc/certs/server.pem",
        "interNodeTlsPk" : "security/grpc/keys/server.key.pem",
        "interNodeTlsPkPwd" : "security/grpc/pass/mindie_server_key_pwd.txt",
        "interNodeTlsCrlPath" : "security/grpc/certs/",
        "interNodeTlsCrlFiles" : ["server_crl.pem"],
        "interNodeKmcKsfMaster" : "tools/pmt/master/ksfa",
        "interNodeKmcKsfStandby" : "tools/pmt/standby/ksfb",
        "kvPoolConfig" : {"backend":"", "configPath":""},
        "ModelDeployConfig" :
        {
            "maxSeqLen" : 131072,
            "maxInputTokenLen" : 131072,
            "truncation" : false,
            "ModelConfig" : [
                {
                    "modelInstanceType" : "Standard",
                    "modelName" : "ds3.1t-w8a8c8",
                    "modelWeightPath" : "/ai/DeepSeek-V3.1-Terminus-w8a8c8/",
                    "worldSize" : 8,
                    "cpuMemSize" : 5,
                    "npuMemSize" : -1,
                    "backendType" : "atb",
                    "trustRemoteCode" : false,
                    "async_scheduler_wait_time": 120,
                    "kv_trans_timeout": 10,
                    "kv_link_timeout": 1080,
                    "dp": 1,
                    "sp": 8,
                    "tp": 8,
                    "cp": 2,
                    "moe_ep": 16,
                    "moe_tp": 1,
                    "models": {
                        "deepseekv2": {
                            "ep_level": 1,
                            "enable_init_routing_cutoff": true,
                            "topk_scaling_factor": 0.25,
                            "enable_oproj_prefetch": true,
                            "enable_mlapo_prefetch": true,
                            "kv_cache_options": {"enable_nz": true},
			                "tool_call_options": {"tool_call_parser": "deepseekv31"},
                            "chat_template": "/ai/DeepSeek-V3.1-Terminus-w8a8c8/tool_chat_template_deepseekv31.jinja"
                       }
                   },
                   "llm": {
                        "parallel_options": {
                        "dense_mlp_local_tp": 16
                        }
                }
                }
            ]
        },

        "ScheduleConfig" :
        {
            "templateType" : "Standard",
            "templateName" : "Standard_LLM",
            "cacheBlockSize" : 128,

            "maxPrefillBatchSize" : 10,
            "maxPrefillTokens" : 131072,
            "prefillTimeMsPerReq" : 150,
            "prefillPolicyType" : 0,

            "decodeTimeMsPerReq" : 50,
            "decodePolicyType" : 0,

            "maxBatchSize" : 20,
            "maxIterTimes" : 131072,
            "maxPreemptCount" : 0,
            "supportSelectBatch" : false,
            "maxQueueDelayMicroseconds" : 5000
        }
    },

    "LogConfig": {
        "dynamicLogLevel" : "",
        "dynamicLogLevelValidHours" : 2,
        "dynamicLogLevelValidTime" : ""
    }
}

主节点IP和环境变量env.sh里面配置的MASTER_IP的值一致。 保存配置文件后就可以拉起服务化了。

3.4.4 拉起服务化推理

使用下面命令拉起服务：

cd /usr/local/Ascend/mindie/latest/mindie-service/
nohup ./bin/mindieservice_daemon > output.log 2>&1 &

查看日志命令

tail -f output.log

等待5~10分钟，直到日志中显示“Daemon start success!”表示推理服务拉起成功。

3.5 推理接口测试

在宿主机上面执行curl命令测试接口是否正常返回。

time curl -H "Accept: application/json" -H "Content-type: application/json" -X POST -d '{
"model": "ds3.1t-w8a8c8",
"messages": [
{
"role": "user",
"content": "中国有哪些景点？"
}
],
"max_tokens": 300,
"temperature": 0.5,
"top_p": 0.96,
"stream": false
}' http://{主节点IP}:1025/v1/chat/completions

4. 性能测试

使用Run_Benchmark工具测试模型推理性能，工具介绍：https://gitcode.com/Ascend-SACT/Run_Benchmark

此工具基于AISBench评测工具开发，集成了文本、图片、视频、音频的数据集，可以方便地对各种模型进行评测。

4.1 安装AISBench

运行性能测试工具Run_Benchmark之前需要安装好AISBench，安装指导参考https://gitee.com/aisbench/benchmark#-%E5%B7%A5%E5%85%B7%E5%AE%89%E8%A3%85。

安装完成后执行ais_bench -h命令，如果正常回显命令参数则说明安装成功。

4.2 下载并配置Run_Benchmark

打开Run_Benchmark工具首页，下载zip格式源代码。上传到服务器并做如下配置。

vi run_benchmark.sh

对下面的测试配置做修改：

VERSION=""       
Sever_NAME=""   #服务化时使用的model-name
SERVICE_IP=""   #本地推理服务器的IP
SERVICE_PORT=""    #拉起服务时设置的端口
MODEL_PATH=""      #权重路径
#8卡PD混部，下面2个参数都配置为8
D_NUM=8
ALL_NUM=8
#相当于轮次。测试请求总数=并发数*轮次
Concurrent_Multiplier=4
#测试DeepSeek的NLP模型，使用SYN合成数据集即可：
PARAM_SETS=(
  "0 1 32768 1024 SYN 0"
  "0 2 32768 1024 SYN 0"
  "0 4 32768 1024 SYN 0"
  "0 8 32768 1024 SYN 0"
)
#上面的例子中表示分别跑4个并发测试任务，任务的并发数分别为1/2/4/8，每个任务都使用64K输入，1K输出，SYN类型的测试集，跑4个轮次。
#填写pip show ais-bench-benchmark命令中显示的Location路径
AISBENCHMARK_PATH=""
#搜索Data_Command_Type，修改数据集名称
#本案例使用MindIE推理框架，对于SYN类型的数据集，使用synthetic_gen数据集，默认脚本中不用修改。

保存修改后的文件。

4.3 运行测试任务

由于测试请求数量越多，测试耗时就越长，因此采用nohup方式启动测试任务：

nohup bash run_benchmark.sh > test_ais_bench.log &

等待测试任务运行结束后，在目标目录下查看summary.csv内容即为测试结果。

当前版本每次运行完性能测试脚本后，都会在summary.csv内容末尾追加测试结果，测试结果中包含了TTFT、TPOP、吞吐量、输出吞吐量、单卡吞吐量、QPS、QPM等指标，可以根据测试结果来调整业务参数获得最佳性能。

5. 常见问题

5.1 拉起服务报错“RuntimeError: NPU out of memory”

显存不足，一般是MindIE的config.json文件中参数maxSeqLen、maxInputTokenLen、maxPrefillTokens、maxIterTimes配置不合理导致的，参考官方参数解释说明进行排查。