昇腾实战派
更多内容可查看【DeepSeek推理部署】知识地图

1. 模型概述及场景

DeepSeek-V3.2是一款在高计算效率与卓越推理能力间取得突破性平衡的先进模型，基于以下关键技术突破：

DeepSeek 稀疏注意力（DSA）： 引入DSA高效的注意力机制，它显著降低了计算复杂性，同时保持了模型性能，特别针对长上下文场景进行了优化。

可扩展的强化学习框架： 通过实施强大的 RL 协议并扩展后训练计算，DeepSeek-V3.2 的表现与 GPT-5 相当。

w8a8权重下载链接：https://modelers.cn/models/Eco-Tech/DeepSeek-V3.2-w8a8-QuaRot/

2. 准备运行环境

2.1 硬件版本

组件	版本
硬件环境	Atlas 800T A2（16卡）

2.2 软件版本

组件	版本
MindIE	2.2.T32
HDK	Ascend HDK 25.2.1
CANN	8.3.RC1
模型	DeepSeek V3.2

3. 运行指导

3.1 获取镜像

镜像获取链接：https://www.hiascend.com/developer/ascendhub/detail/af85b724a7e5469ebd7ea13c3439d48f

3.2 创建容器

docker run -itd --privileged  --name=mindie-test --net=host \
--shm-size 500g \
--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 /disk1:/disk1 \
-v /disk2:/disk2 \
-v /disk3:/disk3 \
-v /opt:/opt \
-v /home:/home \
--entrypoint /bin/bash mindiexxx

3.3 生成ranktable表

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

3.4 修改服务化参数

供参考：

3.4.1 主节点：

{
    "BackendConfig": {
        "ModelDeployConfig": {
            "ModelConfig": [
                {
                    "backendType": "atb",
                    "cpuMemSize": 5,
                    "dp": 1,
                    "enable_warmup_with_sampling": false,
                    "ignore_eos": true,
                    "modelInstanceType": "Standard",
                    "modelName": "m_model",
                    "modelWeightPath": "/disk2/models/dsv3.2",
                    "models": {
                        "deepseekv2": {
                            "enable_mlapo_prefetch": true,
                            "kv_cache_options": {
                                "enable_nz": true
                            }
                        }
                    },
                    "moe_ep": 1,
                    "moe_tp": 16,
                    "npuMemSize": 4,
                    "plugin_params": "{\"plugin_type\":\"mtp\",\"num_speculative_tokens\": 1}",
                    "sp": 1,
                    "tp": 16,
                    "trustRemoteCode": false,
                    "worldSize": 8
                }
            ],
            "maxInputTokenLen": 8384,
            "maxSeqLen": 8384,
            "truncation": false
        },
        "ScheduleConfig": {
            "cacheBlockSize": 128,
            "decodePolicyType": 0,
            "decodeTimeMsPerReq": 50,
            "maxBatchSize": 200,
            "maxIterTimes": 8384,
            "maxPreemptCount": 0,
            "maxPrefillBatchSize": 1,
            "maxPrefillTokens": 8384,
            "maxQueueDelayMicroseconds": 5000,
            "prefillPolicyType": 0,
            "prefillTimeMsPerReq": 150,
            "supportSelectBatch": false,
            "templateName": "Standard_LLM",
            "templateType": "Standard"
        },
        "backendName": "mindieservice_llm_engine",
        "interNodeKmcKsfMaster": "tools/pmt/master/ksfa",
        "interNodeKmcKsfStandby": "tools/pmt/standby/ksfb",
        "interNodeTLSEnabled": false,
        "interNodeTlsCaFiles": [
            "ca.pem"
        ],
        "interNodeTlsCaPath": "security/grpc/ca/",
        "interNodeTlsCert": "security/grpc/certs/server.pem",
        "interNodeTlsCrlFiles": [
            "server_crl.pem"
        ],
        "interNodeTlsCrlPath": "security/grpc/certs/",
        "interNodeTlsPk": "security/grpc/keys/server.key.pem",
        "interNodeTlsPkPwd": "security/grpc/pass/mindie_server_key_pwd.txt",
        "modelInstanceNumber": 1,
        "multiNodesInferEnabled": true,
        "multiNodesInferPort": 1120,
        "npuDeviceIds": [
            [
                0,
                1,
                2,
                3,
                4,
                5,
                6,
                7
            ]
        ],
        "tokenizerProcessNumber": 8
    },
    "LogConfig": {
        "dynamicLogLevel": "",
        "dynamicLogLevelValidHours": 2,
        "dynamicLogLevelValidTime": ""
    },
    "ServerConfig": {
        "allowAllZeroIpListening": false,
        "distDPServerEnabled": false,
        "e2eTimeout": 3600,
        "fullTextEnabled": false,
        "httpsEnabled": false,
        "inferMode": "standard",
        "interCommPk": "security/grpc/keys/server.key.pem",
        "interCommPkPwd": "security/grpc/pass/key_pwd.txt",
        "interCommPort": 1121,
        "interCommTLSEnabled": false,
        "interCommTlsCaFiles": [
            "ca.pem"
        ],
        "interCommTlsCaPath": "security/grpc/ca/",
        "interCommTlsCert": "security/grpc/certs/server.pem",
        "interCommTlsCrlFiles": [
            "server_crl.pem"
        ],
        "interCommTlsCrlPath": "security/grpc/certs/",
        "ipAddress": "此次为MASTER_IP",
        "kmcKsfMaster": "tools/pmt/master/ksfa",
        "kmcKsfStandby": "tools/pmt/standby/ksfb",
        "managementIpAddress": "127.0.0.2",
        "managementPort": 1026,
        "managementTlsCaFile": [
            "management_ca.pem"
        ],
        "managementTlsCert": "security/certs/management/server.pem",
        "managementTlsCrlFiles": [
            "server_crl.pem"
        ],
        "managementTlsCrlPath": "security/management/certs/",
        "managementTlsPk": "security/keys/management/server.key.pem",
        "managementTlsPkPwd": "security/pass/management/key_pwd.txt",
        "maxLinkNum": 1000,
        "metricsPort": 1027,
        "openAiSupport": "vllm",
        "port": 8080,
        "tlsCaFile": [
            "ca.pem"
        ],
        "tlsCaPath": "security/ca/",
        "tlsCert": "security/certs/server.pem",
        "tlsCrlFiles": [
            "server_crl.pem"
        ],
        "tlsCrlPath": "security/certs/",
        "tlsPk": "security/keys/server.key.pem",
        "tlsPkPwd": "security/pass/key_pwd.txt",
        "tokenTimeout": 3600
    },
    "Version": "1.0.0"
}

3.4.2 从节点：

{
    "BackendConfig": {
        "ModelDeployConfig": {
            "ModelConfig": [
                {
                    "backendType": "atb",
                    "cpuMemSize": 5,
                    "dp": 1,
                    "enable_warmup_with_sampling": false,
                    "ignore_eos": true,
                    "modelInstanceType": "Standard",
                    "modelName": "m_model",
                    "modelWeightPath": "/disk2/models/dsv3.2",
                    "models": {
                        "deepseekv2": {
                            "enable_mlapo_prefetch": true,
                            "kv_cache_options": {
                                "enable_nz": true
                            }
                        }
                    },
                    "moe_ep": 1,
                    "moe_tp": 16,
                    "npuMemSize": 4,
                    "plugin_params": "{\"plugin_type\":\"mtp\",\"num_speculative_tokens\": 1}",
                    "sp": 1,
                    "tp": 16,
                    "trustRemoteCode": false,
                    "worldSize": 8
                }
            ],
            "maxInputTokenLen": 8384,
            "maxSeqLen": 8384,
            "truncation": false
        },
        "ScheduleConfig": {
            "cacheBlockSize": 128,
            "decodePolicyType": 0,
            "decodeTimeMsPerReq": 50,
            "maxBatchSize": 200,
            "maxIterTimes": 8384,
            "maxPreemptCount": 0,
            "maxPrefillBatchSize": 1,
            "maxPrefillTokens": 8384,
            "maxQueueDelayMicroseconds": 5000,
            "prefillPolicyType": 0,
            "prefillTimeMsPerReq": 150,
            "supportSelectBatch": false,
            "templateName": "Standard_LLM",
            "templateType": "Standard"
        },
        "backendName": "mindieservice_llm_engine",
        "interNodeKmcKsfMaster": "tools/pmt/master/ksfa",
        "interNodeKmcKsfStandby": "tools/pmt/standby/ksfb",
        "interNodeTLSEnabled": false,
        "interNodeTlsCaFiles": [
            "ca.pem"
        ],
        "interNodeTlsCaPath": "security/grpc/ca/",
        "interNodeTlsCert": "security/grpc/certs/server.pem",
        "interNodeTlsCrlFiles": [
            "server_crl.pem"
        ],
        "interNodeTlsCrlPath": "security/grpc/certs/",
        "interNodeTlsPk": "security/grpc/keys/server.key.pem",
        "interNodeTlsPkPwd": "security/grpc/pass/mindie_server_key_pwd.txt",
        "modelInstanceNumber": 1,
        "multiNodesInferEnabled": true,
        "multiNodesInferPort": 1120,
        "npuDeviceIds": [
            [
                0,
                1,
                2,
                3,
                4,
                5,
                6,
                7
            ]
        ],
        "tokenizerProcessNumber": 8
    },
    "LogConfig": {
        "dynamicLogLevel": "",
        "dynamicLogLevelValidHours": 2,
        "dynamicLogLevelValidTime": ""
    },
    "ServerConfig": {
        "allowAllZeroIpListening": false,
        "distDPServerEnabled": false,
        "e2eTimeout": 3600,
        "fullTextEnabled": false,
        "httpsEnabled": false,
        "inferMode": "standard",
        "interCommPk": "security/grpc/keys/server.key.pem",
        "interCommPkPwd": "security/grpc/pass/key_pwd.txt",
        "interCommPort": 1121,
        "interCommTLSEnabled": false,
        "interCommTlsCaFiles": [
            "ca.pem"
        ],
        "interCommTlsCaPath": "security/grpc/ca/",
        "interCommTlsCert": "security/grpc/certs/server.pem",
        "interCommTlsCrlFiles": [
            "server_crl.pem"
        ],
        "interCommTlsCrlPath": "security/grpc/certs/",
        "ipAddress": "127.0.0.1",
        "kmcKsfMaster": "tools/pmt/master/ksfa",
        "kmcKsfStandby": "tools/pmt/standby/ksfb",
        "managementIpAddress": "127.0.0.2",
        "managementPort": 1026,
        "managementTlsCaFile": [
            "management_ca.pem"
        ],
        "managementTlsCert": "security/certs/management/server.pem",
        "managementTlsCrlFiles": [
            "server_crl.pem"
        ],
        "managementTlsCrlPath": "security/management/certs/",
        "managementTlsPk": "security/keys/management/server.key.pem",
        "managementTlsPkPwd": "security/pass/management/key_pwd.txt",
        "maxLinkNum": 1000,
        "metricsPort": 1027,
        "openAiSupport": "vllm",
        "port": 8080,
        "tlsCaFile": [
            "ca.pem"
        ],
        "tlsCaPath": "security/ca/",
        "tlsCert": "security/certs/server.pem",
        "tlsCrlFiles": [
            "server_crl.pem"
        ],
        "tlsCrlPath": "security/certs/",
        "tlsPk": "security/keys/server.key.pem",
        "tlsPkPwd": "security/pass/key_pwd.txt",
        "tokenTimeout": 3600
    },
    "Version": "1.0.0"
}

3.3 部署服务

3.3.1 修改模型config文件

修改模型配置文件config.json，将model_type改为deepseek_v3：

"model_type": "deepseek_v3"`

3.3.2 主节点启动

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
source /usr/local/Ascend/mindie/set_env.sh

export MIES_CONTAINER_IP=xxx1
# ranktable.json文件位置
export RANKTABLEFILE=xxx/hccl_2s_16p.json
# 主节点 ip 和端口
export MASTER_IP=xxx1
export MASTER_PORT=8088
# 使能内存池扩展段特性
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
# 提升权重加载速度
export OMP_NUM_THREADS=10

export HCCL_CONNECT_TIMEOUT=7200
export HCCL_HOST_SOCKET_PORT_RANGE=60000-60050
export MINDIE_LOG_LEVEL=INFO

cd /usr/local/Ascend/mindie/latest/mindie-service/
./bin/mindieservice_daemon

3.3.2 从节点启动

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
source /usr/local/Ascend/mindie/set_env.sh

export MIES_CONTAINER_IP=xxx2
# ranktable.json文件位置
export RANKTABLEFILE=xxx/hccl_2s_16p.json
# 主节点 ip 和端口
export MASTER_IP=xxx1
export MASTER_PORT=8088
# 使能内存池扩展段特性
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
# 提升权重加载速度
export OMP_NUM_THREADS=10

export HCCL_CONNECT_TIMEOUT=7200
export HCCL_HOST_SOCKET_PORT_RANGE=60000-60050
export MINDIE_LOG_LEVEL=INFO

cd /usr/local/Ascend/mindie/latest/mindie-service/
./bin/mindieservice_daemon

注意：MASTER_PORT不要和config.json内的port冲突。

3.4 发送curl请求验证

curl -X POST http://xxx:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
    "model": "m_model",
    "messages": [
        {
            "role": "user",
            "content": "你是谁？"
        }
    ],
    "max_tokens": 100,
    "ignore_eos": false,
    "stream": false
}'

curl -X POST http://xxx:8080/v1/completions   -H "Content-Type: application/json"   -d '{
    "model": "m_model",
    "prompt": "<｜begin▁of▁sentence｜><｜User｜>hello<｜Assistant｜></think>Hello! I am DeepSeek.<｜end▁of▁sentence｜><｜User｜>自我介绍<｜Assistant｜><think>",
    "max_tokens": 100,
    "temperature": 1.0,
    "top_p": 0.95,
    "stop": ["<｜end▁of▁sentence｜>"]
  }'

4. 问题记录

4.1 问题描述：The model does not appear to be a chat model because it is not configured with a chat_template.

解决方案：
需要修改tokenizer_config.json，DeepSeek-V3.2内的tokenizer_config.json缺少chat_templat，可以采用DeepSeek-V3.2-Exp内的chat_template。

{
  "add_bos_token": false,
  "add_eos_token": false,
  "bos_token": {
    "__type": "AddedToken",
    "content": "<｜begin▁of▁sentence｜>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  },
  "clean_up_tokenization_spaces": false,
  "eos_token": {
    "__type": "AddedToken",
    "content": "<｜end▁of▁sentence｜>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  },
  "legacy": true,
  "model_max_length": 131072,
  "pad_token": {
    "__type": "AddedToken",
    "content": "<｜end▁of▁sentence｜>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  },
  "sp_model_kwargs": {},
  "unk_token": null,
  "tokenizer_class": "LlamaTokenizerFast",
  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% if not thinking is defined %}{% set thinking = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, system_prompt='', is_first_sp=true, is_last_user=false, is_only_sys=false, is_prefix=false) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + message['content'] %}{%- endif %}{% set ns.is_only_sys = true %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{%- set ns.is_first = false -%}{%- set ns.is_last_user = true -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['tool_calls'] is defined and message['tool_calls'] is not none %}{%- if ns.is_last_user or ns.is_only_sys %}{{'<｜Assistant｜></think>'}}{%- endif %}{%- set ns.is_last_user = false -%}{%- set ns.is_first = false %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>'+ tool['function']['name'] + '<｜tool▁sep｜>' + tool['function']['arguments'] + '<｜tool▁call▁end｜>'}}{%- else %}{{message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['function']['name'] + '<｜tool▁sep｜>' + tool['function']['arguments'] + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'<｜tool▁call▁begin｜>'+ tool['function']['name'] + '<｜tool▁sep｜>' + tool['function']['arguments'] + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and (message['tool_calls'] is not defined or message['tool_calls'] is none) %}{%- if ns.is_last_user %}{{'<｜Assistant｜>'}}{%- if message['prefix'] is defined and message['prefix'] and thinking %}{{'<think>'}}{%- else %}{{'</think>'}}{%- endif %}{%- endif %}{%- if message['prefix'] is defined and message['prefix'] %}{%- set ns.is_prefix = true -%}{%- endif %}{%- set ns.is_last_user = false -%}{%- if ns.is_tool %}{{message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{%- set content = message['content'] -%}{%- if '</think>' in content %}{%- set content = content.split('</think>', 1)[1] -%}{%- endif %}{{content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_last_user = false -%}{%- set ns.is_tool = true -%}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- if message['role'] != 'system' %}{% set ns.is_only_sys = false %}{%- endif %}{%- endfor -%}{% if add_generation_prompt and not ns.is_tool%}{% if ns.is_last_user or ns.is_only_sys or not ns.is_prefix %}{{'<｜Assistant｜>'}}{%- if not thinking %}{{'</think>'}}{%- else %}{{'<think>'}}{%- endif %}{% endif %}{% endif %}"
}