[{"data":1,"prerenderedAt":876},["ShallowReactive",2],{"content-query-EnGmRKWWFE":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":870,"_id":871,"_source":872,"_file":873,"_stem":874,"_extension":875},"/technology-blogs/zh/3723","zh",false,"","基于昇思MindSpore，中国电信发布TeleChat系列4个模型并开源","5月12日，中国电信开源TeleChat系列四个模型，涵盖复杂推理和通用问答的多个尺寸模型，包括TeleChat-T1-35B、TeleChat-T1-115B、TeleChat2.5-35B和TeleChat2.5-115B，实测模型性能均有显著的性能效果。","2025-05-12","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/05/16/c18c2a27297a4d60bb2965254646042c.png","technology-blogs","实践",{"type":15,"children":16,"toc":864},"root",[17,25,31,36,47,56,65,74,83,91,99,107,138,146,151,166,174,182,190,198,235,242,251,259,267,280,288,293,311,318,325,333,341,349,354,361,369,377,385,390,398,403,413,427,435,440,448,453,466,473,481,486,494,502,510,515,523,528,546,551,559,567,577,584,592,600,604,612,617,625,630,635,643,650,658,663,671,679,712,720,728,736,744,753,760,767,774,778,786,794,802,836,843,851,859],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"基于昇思mindspore中国电信发布telechat系列4个模型并开源",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":24,"value":30},"5月12日，中国电信开源TeleChat系列四个模型，涵盖复杂推理和通用问答的多个尺寸模型，包括TeleChat-T1-35B、TeleChat-T1-115B、TeleChat2.5-35B和TeleChat2.5-115B，实测模型性能均有显著的性能效果。TeleChat系列模型基于昇思MindSpore+vLLM打造，现已上线魔乐社区，欢迎广大开发者下载体验！",{"type":18,"tag":26,"props":32,"children":33},{},[34],{"type":24,"value":35},"开源链接：",{"type":18,"tag":26,"props":37,"children":38},{},[39],{"type":18,"tag":40,"props":41,"children":45},"a",{"href":42,"rel":43},"https://modelers.cn/models/MindSpore-Lab/T1-35B",[44],"nofollow",[46],{"type":24,"value":42},{"type":18,"tag":26,"props":48,"children":49},{},[50],{"type":18,"tag":40,"props":51,"children":54},{"href":52,"rel":53},"https://modelers.cn/models/MindSpore-Lab/T1-115B",[44],[55],{"type":24,"value":52},{"type":18,"tag":26,"props":57,"children":58},{},[59],{"type":18,"tag":40,"props":60,"children":63},{"href":61,"rel":62},"https://modelers.cn/models/MindSpore-Lab/TeleChat2.5-35B",[44],[64],{"type":24,"value":61},{"type":18,"tag":26,"props":66,"children":67},{},[68],{"type":18,"tag":40,"props":69,"children":72},{"href":70,"rel":71},"https://modelers.cn/models/MindSpore-Lab/TeleChat2.5-115B",[44],[73],{"type":24,"value":70},{"type":18,"tag":26,"props":75,"children":76},{},[77],{"type":18,"tag":78,"props":79,"children":80},"strong",{},[81],{"type":24,"value":82},"# 01",{"type":18,"tag":26,"props":84,"children":85},{},[86],{"type":18,"tag":78,"props":87,"children":88},{},[89],{"type":24,"value":90},"模型介绍",{"type":18,"tag":26,"props":92,"children":93},{},[94],{"type":18,"tag":78,"props":95,"children":96},{},[97],{"type":24,"value":98},"1",{"type":18,"tag":26,"props":100,"children":101},{},[102],{"type":18,"tag":78,"props":103,"children":104},{},[105],{"type":24,"value":106},"TeleChat T1模型",{"type":18,"tag":26,"props":108,"children":109},{},[110,115,117,122,124,129,131,136],{"type":18,"tag":78,"props":111,"children":112},{},[113],{"type":24,"value":114},"T1",{"type":24,"value":116}," 模型是 ",{"type":18,"tag":78,"props":118,"children":119},{},[120],{"type":24,"value":121},"TeleChat",{"type":24,"value":123}," 系列专注于复杂推理的模型，由中国电信人工智能研究院研发训练。该系列模型借助先进的思维推理和批判纠错能力，在下游复杂任务中有很好的表现。本次中国电信开源了 ",{"type":18,"tag":78,"props":125,"children":126},{},[127],{"type":24,"value":128},"T1-35B",{"type":24,"value":130}," 和 ",{"type":18,"tag":78,"props":132,"children":133},{},[134],{"type":24,"value":135},"T1-115B",{"type":24,"value":137}," 两款不同尺寸的模型，与同尺寸模型相比都具有较好的效果表现。",{"type":18,"tag":26,"props":139,"children":140},{},[141],{"type":18,"tag":78,"props":142,"children":143},{},[144],{"type":24,"value":145},"训练策略",{"type":18,"tag":26,"props":147,"children":148},{},[149],{"type":24,"value":150},"采用课程学习贯穿全流程的后训练方案，循序渐进提升模型效果。",{"type":18,"tag":152,"props":153,"children":154},"ul",{},[155,161],{"type":18,"tag":156,"props":157,"children":158},"li",{},[159],{"type":24,"value":160},"微调阶段：将多任务数据集进行难度划分（根据模型推理正误比率判断），首先使用中低难度冷启动微调，然后使用RFT方式筛选中高难度数据进行持续微调进行效果提升；",{"type":18,"tag":156,"props":162,"children":163},{},[164],{"type":24,"value":165},"强化学习阶段：首先对数理逻辑、代码能力进行提升，采用难度渐进式课程学习方案进行能力强化；然后，基于指令遵循、安全、幻觉、Function Call等10多种混合通用任务进行持续强化，全面提升模型效果。",{"type":18,"tag":26,"props":167,"children":168},{},[169],{"type":18,"tag":78,"props":170,"children":171},{},[172],{"type":24,"value":173},"效果评测",{"type":18,"tag":26,"props":175,"children":176},{},[177],{"type":18,"tag":178,"props":179,"children":181},"img",{"alt":7,"src":180},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/05/16/af0ac90a77364272908696836535200a.png",[],{"type":18,"tag":26,"props":183,"children":184},{},[185],{"type":18,"tag":78,"props":186,"children":187},{},[188],{"type":24,"value":189},"2",{"type":18,"tag":26,"props":191,"children":192},{},[193],{"type":18,"tag":78,"props":194,"children":195},{},[196],{"type":24,"value":197},"TeleChat2.5模型",{"type":18,"tag":26,"props":199,"children":200},{},[201,206,208,212,214,219,221,226,228],{"type":18,"tag":78,"props":202,"children":203},{},[204],{"type":24,"value":205},"TeleChat2.5",{"type":24,"value":207}," 是 ",{"type":18,"tag":78,"props":209,"children":210},{},[211],{"type":24,"value":121},{"type":24,"value":213}," 系列新版通用问答模型，由中国电信人工智能研究院**（TeleAI）**研发训练，包括了 ",{"type":18,"tag":78,"props":215,"children":216},{},[217],{"type":24,"value":218},"TeleChat2.5-35B",{"type":24,"value":220}," 与 ",{"type":18,"tag":78,"props":222,"children":223},{},[224],{"type":24,"value":225},"TeleChat2.5-115B",{"type":24,"value":227},"。TeleChat2.5 基于最新强化的 TeleBase2.5 系列模型进行训练，在理科、通用问答、Function Call等任务上有显著的效果提升。TeleChat2.5 的微调方法延续了 TeleChat2 系列，具体请参考 TeleChat2（",{"type":18,"tag":40,"props":229,"children":232},{"href":230,"rel":231},"https://github.com/Tele-AI/TeleChat2%EF%BC%89%E3%80%82",[44],[233],{"type":24,"value":234},"https://github.com/Tele-AI/TeleChat2）。",{"type":18,"tag":26,"props":236,"children":237},{},[238],{"type":18,"tag":78,"props":239,"children":240},{},[241],{"type":24,"value":145},{"type":18,"tag":243,"props":244,"children":246},"h4",{"id":245},"数据",[247],{"type":18,"tag":78,"props":248,"children":249},{},[250],{"type":24,"value":245},{"type":18,"tag":152,"props":252,"children":253},{},[254],{"type":18,"tag":156,"props":255,"children":256},{},[257],{"type":24,"value":258},"为了提高模型训练数据的数量和质量，TeleChat2.5 在训练过程中采用了大量理科学科和编程领域的合成数据。在合成过程中，为了减少错误信息的引入，主要以基于知识点或知识片段的教育类知识合成为主。",{"type":18,"tag":243,"props":260,"children":262},{"id":261},"基础模型训练",[263],{"type":18,"tag":78,"props":264,"children":265},{},[266],{"type":24,"value":261},{"type":18,"tag":152,"props":268,"children":269},{},[270,275],{"type":18,"tag":156,"props":271,"children":272},{},[273],{"type":24,"value":274},"TeleChat2.5 采用了多阶段课程学习策略，在训练过程中逐步提升理科和编程类高密度知识数据的比例。每个训练阶段都使用比前一阶段质量更高、难度更大的数据，以实现持续的模型优化。",{"type":18,"tag":156,"props":276,"children":277},{},[278],{"type":24,"value":279},"在最终训练阶段，为了平衡模型在各个维度的能力表现，选取了不同训练阶段效果较优的多个模型，并基于各模型的综合表现进行参数加权融合，其中权重分配与模型性能呈正相关。",{"type":18,"tag":243,"props":281,"children":283},{"id":282},"后训练阶段",[284],{"type":18,"tag":78,"props":285,"children":286},{},[287],{"type":24,"value":282},{"type":18,"tag":26,"props":289,"children":290},{},[291],{"type":24,"value":292},"TeleChat2.5采用分阶段优化的模型训练策略：",{"type":18,"tag":152,"props":294,"children":295},{},[296,301,306],{"type":18,"tag":156,"props":297,"children":298},{},[299],{"type":24,"value":300},"融合优化阶段：整合复杂推理与通用问答能力，针对语言理解、数理逻辑等薄弱任务进行解构重组。通过重构任务框架并融合多维度解题思路，生成优化后的通用答案集。此阶段答案长度会适度增加，并基于优化数据实施微调训练。",{"type":18,"tag":156,"props":302,"children":303},{},[304],{"type":24,"value":305},"能力强化阶段：针对数理逻辑与编程类任务，通过注入结构化解题思路，结合基于规则的强化学习奖励机制，显著提升模型对复杂任务的理解与处理能力。",{"type":18,"tag":156,"props":307,"children":308},{},[309],{"type":24,"value":310},"泛化提升阶段：面向安全合规、指令响应、函数调用、数学推理、代码生成等十余种任务类型进行系统性强化学习增强，全面提升模型的通用任务处理能力。",{"type":18,"tag":26,"props":312,"children":313},{},[314],{"type":18,"tag":78,"props":315,"children":316},{},[317],{"type":24,"value":173},{"type":18,"tag":26,"props":319,"children":320},{},[321],{"type":18,"tag":178,"props":322,"children":324},{"alt":7,"src":323},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/05/16/4cbbebeeaf26430c8e9488b1f998536a.png",[],{"type":18,"tag":26,"props":326,"children":327},{},[328],{"type":18,"tag":78,"props":329,"children":330},{},[331],{"type":24,"value":332},"以下为手把手教程：（以T1-35B模型为例）",{"type":18,"tag":26,"props":334,"children":335},{},[336],{"type":18,"tag":78,"props":337,"children":338},{},[339],{"type":24,"value":340},"# 02",{"type":18,"tag":26,"props":342,"children":343},{},[344],{"type":18,"tag":78,"props":345,"children":346},{},[347],{"type":24,"value":348},"快速开始",{"type":18,"tag":26,"props":350,"children":351},{},[352],{"type":24,"value":353},"T1-35B 推理至少需要1台（2卡）Atlas 800T A2（64G显存规格）服务器。昇思MindSpore提供了T1-35B 推理可用的Docker容器镜像，供开发者快速体验。",{"type":18,"tag":26,"props":355,"children":356},{},[357],{"type":18,"tag":78,"props":358,"children":359},{},[360],{"type":24,"value":98},{"type":18,"tag":26,"props":362,"children":363},{},[364],{"type":18,"tag":78,"props":365,"children":366},{},[367],{"type":24,"value":368},"模型下载",{"type":18,"tag":26,"props":370,"children":371},{},[372],{"type":18,"tag":78,"props":373,"children":374},{},[375],{"type":24,"value":376},"社区",{"type":18,"tag":26,"props":378,"children":379},{},[380],{"type":18,"tag":78,"props":381,"children":382},{},[383],{"type":24,"value":384},"下载地址",{"type":18,"tag":26,"props":386,"children":387},{},[388],{"type":24,"value":389},"魔乐社区",{"type":18,"tag":26,"props":391,"children":392},{},[393],{"type":18,"tag":40,"props":394,"children":396},{"href":42,"rel":395},[44],[397],{"type":24,"value":42},{"type":18,"tag":26,"props":399,"children":400},{},[401],{"type":24,"value":402},"执行以下命令为自定义下载路径/home/teleAI/T1-35B添加白名单：",{"type":18,"tag":404,"props":405,"children":407},"pre",{"code":406},"export HUB_WHITE_LIST_PATHS=/home/teleAI/T1-35B\n",[408],{"type":18,"tag":409,"props":410,"children":411},"code",{"__ignoreMap":7},[412],{"type":24,"value":406},{"type":18,"tag":26,"props":414,"children":415},{},[416,418,425],{"type":24,"value":417},"执行以下 Python 脚本从魔乐社区下载昇思 MindSpore 版本的 T1-35B 文件至指定路径 /home/teleAI/T1-35B （需提前pip安装openmind_hub库，相关参考文档可参考：openMind Hub Client使用教程（",{"type":18,"tag":40,"props":419,"children":422},{"href":420,"rel":421},"https://modelers.cn/docs/zh/openmind-hub-client/0.9/overview.html%EF%BC%89%EF%BC%89%E3%80%82%E4%B8%8B%E8%BD%BD%E7%9A%84%E6%96%87%E4%BB%B6%E5%8C%85%E5%90%AB%E6%A8%A1%E5%9E%8B%E4%BB%A3%E7%A0%81%E3%80%81%E6%9D%83%E9%87%8D%E3%80%81%E5%88%86%E8%AF%8D%E6%A8%A1%E5%9E%8B%E5%92%8C%E7%A4%BA%E4%BE%8B%E4%BB%A3%E7%A0%81%EF%BC%8C%E5%8D%A0%E7%94%A8%E7%BA%A6",[44],[423],{"type":24,"value":424},"https://modelers.cn/docs/zh/openmind-hub-client/0.9/overview.html））。下载的文件包含模型代码、权重、分词模型和示例代码，占用约",{"type":24,"value":426}," 67GB 的磁盘空间：",{"type":18,"tag":404,"props":428,"children":430},{"code":429},"\nfrom openmind_hub import snapshot_download\n \nsnapshot_download(\n    repo_id=\"MindSpore-Lab/T1-35B\",\n    local_dir=\"/home/teleAI/T1-35B\",\n    local_dir_use_symlinks=False,\n)\n",[431],{"type":18,"tag":409,"props":432,"children":433},{"__ignoreMap":7},[434],{"type":24,"value":429},{"type":18,"tag":26,"props":436,"children":437},{},[438],{"type":24,"value":439},"下载完成的模型文件夹目录结构如下：",{"type":18,"tag":404,"props":441,"children":443},{"code":442},"\nT1-35B/\n    ├── config.json                               # 模型json配置文件\n    ├── configuration_telechat2.py\n    ├── generation_config.json\n    ├── generation_utils.py\n    ├── modeling_telechat2.py\n    ├── special_tokens_map.json\n    ├── tokenization_telechat2.py\n    ├── tokenizer.model                           # 词表model文件\n    ├── tokenizer_config.json                     # 词表配置文件\n    ├── model-000xx-of-000xx.safetensors          # 模型权重文件\n    └── model.safetensors.index.json              # 模型权重映射文件\n",[444],{"type":18,"tag":409,"props":445,"children":446},{"__ignoreMap":7},[447],{"type":24,"value":442},{"type":18,"tag":26,"props":449,"children":450},{},[451],{"type":24,"value":452},"注意事项：",{"type":18,"tag":152,"props":454,"children":455},{},[456,461],{"type":18,"tag":156,"props":457,"children":458},{},[459],{"type":24,"value":460},"/home/teleAI/T1-35B 可修改为容器可访问的自定义路径，确保该路径有足够的磁盘空间（约 67GB）。",{"type":18,"tag":156,"props":462,"children":463},{},[464],{"type":24,"value":465},"下载时间可能因网络环境而异，建议在稳定的网络环境下操作。",{"type":18,"tag":26,"props":467,"children":468},{},[469],{"type":18,"tag":78,"props":470,"children":471},{},[472],{"type":24,"value":189},{"type":18,"tag":26,"props":474,"children":475},{},[476],{"type":18,"tag":78,"props":477,"children":478},{},[479],{"type":24,"value":480},"下载昇思 MindSpore 推理容器镜像",{"type":18,"tag":26,"props":482,"children":483},{},[484],{"type":24,"value":485},"执行以下 Shell 命令，拉取昇思 MindSpore T1-35B 推理容器镜像：",{"type":18,"tag":404,"props":487,"children":489},{"code":488},"\ndocker pull swr.cn-central-221.ovaijisuan.com/mindformers/mindspore_telechat_t1:20250430\n",[490],{"type":18,"tag":409,"props":491,"children":492},{"__ignoreMap":7},[493],{"type":24,"value":488},{"type":18,"tag":26,"props":495,"children":496},{},[497],{"type":18,"tag":78,"props":498,"children":499},{},[500],{"type":24,"value":501},"3",{"type":18,"tag":26,"props":503,"children":504},{},[505],{"type":18,"tag":78,"props":506,"children":507},{},[508],{"type":24,"value":509},"启动容器",{"type":18,"tag":26,"props":511,"children":512},{},[513],{"type":24,"value":514},"执行以下命令创建并启动容器：",{"type":18,"tag":404,"props":516,"children":518},{"code":517},"docker run -it -u 0 --ipc=host  --network host \\\n--name vllm_telechat_t1 \\\n--privileged \\\n--device=/dev/davinci0 \\\n--device=/dev/davinci1 \\\n--device=/dev/davinci2 \\\n--device=/dev/davinci3 \\\n--device=/dev/davinci4 \\\n--device=/dev/davinci5 \\\n--device=/dev/davinci6 \\\n--device=/dev/davinci7 \\\n--device=/dev/davinci_manager \\\n--device=/dev/devmm_svm \\\n--device=/dev/hisi_hdc \\\n-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \\\n-v /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/ \\\n-v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \\\n-v /usr/local/sbin/:/usr/local/sbin/ \\\n-v /var/log/npu/conf/slog/slog.conf:/var/log/npu/conf/slog/slog.conf \\\n-v /var/log/npu/slog/:/var/log/npu/slog \\\n-v /var/log/npu/profiling/:/var/log/npu/profiling \\\n-v /var/log/npu/dump/:/var/log/npu/dump \\\n-v /var/log/npu/:/usr/slog \\\n-v /home/:/home \\\nswr.cn-central-221.ovaijisuan.com/mindformers/mindspore_telechat_t1:20250430 \\\n/bin/bash\n",[519],{"type":18,"tag":409,"props":520,"children":521},{"__ignoreMap":7},[522],{"type":24,"value":517},{"type":18,"tag":26,"props":524,"children":525},{},[526],{"type":24,"value":527},"注意：",{"type":18,"tag":152,"props":529,"children":530},{},[531,536,541],{"type":18,"tag":156,"props":532,"children":533},{},[534],{"type":24,"value":535},"若模型下载目录不在 /home/ 路径下，需要在 docker run 命令中映射相关的路径（修改-v /home/:/home \\），使容器可以访问对应路径中的文件；",{"type":18,"tag":156,"props":537,"children":538},{},[539],{"type":24,"value":540},"启动容器前，通过 npu-smi info 查看服务器状态，确保卡资源未被其他进程占用。",{"type":18,"tag":156,"props":542,"children":543},{},[544],{"type":24,"value":545},"后续所有操作均在容器中进行。",{"type":18,"tag":26,"props":547,"children":548},{},[549],{"type":24,"value":550},"本文提供两种推理使用方式：服务化部署及离线推理，用户可按需使用。",{"type":18,"tag":26,"props":552,"children":553},{},[554],{"type":18,"tag":78,"props":555,"children":556},{},[557],{"type":24,"value":558},"# 03",{"type":18,"tag":26,"props":560,"children":561},{},[562],{"type":18,"tag":78,"props":563,"children":564},{},[565],{"type":24,"value":566},"服务化部署",{"type":18,"tag":26,"props":568,"children":569},{},[570,575],{"type":18,"tag":78,"props":571,"children":572},{},[573],{"type":24,"value":574},"说明：",{"type":24,"value":576}," 在服务化部署模式下，启动一次vLLM推理服务后，可重复多次发送推理请求。",{"type":18,"tag":26,"props":578,"children":579},{},[580],{"type":18,"tag":78,"props":581,"children":582},{},[583],{"type":24,"value":98},{"type":18,"tag":26,"props":585,"children":586},{},[587],{"type":18,"tag":78,"props":588,"children":589},{},[590],{"type":24,"value":591},"配置环境变量：",{"type":18,"tag":404,"props":593,"children":595},{"code":594},"export MINDFORMERS_MODEL_CONFIG=/usr/local/Python-3.11/lib/python3.11/site-packages/research/telechat2/predict_telechat_35b.yaml\n",[596],{"type":18,"tag":409,"props":597,"children":598},{"__ignoreMap":7},[599],{"type":24,"value":594},{"type":18,"tag":26,"props":601,"children":602},{},[603],{"type":24,"value":189},{"type":18,"tag":26,"props":605,"children":606},{},[607],{"type":18,"tag":78,"props":608,"children":609},{},[610],{"type":24,"value":611},"一键启动vLLM推理",{"type":18,"tag":26,"props":613,"children":614},{},[615],{"type":24,"value":616},"在服务器上执行以下命令启动vLLM推理服务：",{"type":18,"tag":404,"props":618,"children":620},{"code":619},"python3 -m vllm_mindspore.entrypoints vllm.entrypoints.openai.api_server --model \"/home/teleAI/T1-35B\" --port=8000 --trust_remote_code --tensor_parallel_size=2 --max-num-seqs=256 --max_model_len=8192 --max-num-batched-tokens=8192 --block-size=32 --gpu-memory-utilization=0.93\n",[621],{"type":18,"tag":409,"props":622,"children":623},{"__ignoreMap":7},[624],{"type":24,"value":619},{"type":18,"tag":26,"props":626,"children":627},{},[628],{"type":24,"value":629},"注意： 其中/home/teleAI/T1-35B需要修改为实际的模型目录",{"type":18,"tag":26,"props":631,"children":632},{},[633],{"type":24,"value":634},"看到以下日志打印，说明vLLM推理服务启动成功：",{"type":18,"tag":404,"props":636,"children":638},{"code":637},"INFO:     Started server process [xxxxx]\nINFO:     Waiting for application startup.\nINFO:     Application startup complete.\n",[639],{"type":18,"tag":409,"props":640,"children":641},{"__ignoreMap":7},[642],{"type":24,"value":637},{"type":18,"tag":26,"props":644,"children":645},{},[646],{"type":18,"tag":78,"props":647,"children":648},{},[649],{"type":24,"value":501},{"type":18,"tag":26,"props":651,"children":652},{},[653],{"type":18,"tag":78,"props":654,"children":655},{},[656],{"type":24,"value":657},"执行推理请求测试",{"type":18,"tag":26,"props":659,"children":660},{},[661],{"type":24,"value":662},"执行以下命令发送推理请求进行测试：",{"type":18,"tag":404,"props":664,"children":666},{"code":665},"\ncurl http://localhost:8000/v1/completions\n-H \"Content-Type: application/json\" -d '{\n    \"model\": \"/home/teleAI/T1-35B\", \n    \"prompt\": \"\u003C_system>\u003C_user>生抽与老抽的区别？\u003C_bot>\\n\", \n    \"max_tokens\": 2048, \n    \"temperature\": 0.6, \n    \"repetition_penalty\":1.05,\n    \"top_p\":0.95\n}'\n",[667],{"type":18,"tag":409,"props":668,"children":669},{"__ignoreMap":7},[670],{"type":24,"value":665},{"type":18,"tag":26,"props":672,"children":673},{},[674],{"type":18,"tag":78,"props":675,"children":676},{},[677],{"type":24,"value":678},"推理请求报文配置注意事项：",{"type":18,"tag":152,"props":680,"children":681},{},[682,687,692,697,702,707],{"type":18,"tag":156,"props":683,"children":684},{},[685],{"type":24,"value":686},"model: 需要配置为实际的网络权重路径。",{"type":18,"tag":156,"props":688,"children":689},{},[690],{"type":24,"value":691},"prompt: 模板需保持固定，以确保推理时能够生成 reason 过程。可将其中文字部分修改为其他的推理请求问题。",{"type":18,"tag":156,"props":693,"children":694},{},[695],{"type":24,"value":696},"max_tokens: 字段为推理输出最大token长度，可按需修改。",{"type":18,"tag":156,"props":698,"children":699},{},[700],{"type":24,"value":701},"temperature: 建议保持0.6不变",{"type":18,"tag":156,"props":703,"children":704},{},[705],{"type":24,"value":706},"top_p: 建议保持0.95不变",{"type":18,"tag":156,"props":708,"children":709},{},[710],{"type":24,"value":711},"repetition_penalty: 在推理通用任务时，建议使用1.05，可以有效减少重复生成现象；在推理数学、代码任务时，建议使用1.0 。",{"type":18,"tag":26,"props":713,"children":714},{},[715],{"type":18,"tag":78,"props":716,"children":717},{},[718],{"type":24,"value":719},"推理结果：",{"type":18,"tag":404,"props":721,"children":723},{"code":722},"\n嗯，用户问生抽和老抽的区别。首先，我得回忆一下这两种酱油的基本信息。\n \n生抽，我记得是用发酵后的酱油醅提取的，颜色比较浅，味道比较鲜。通常用来炒菜或者凉拌菜，增加鲜味。\n \n老抽的话，应该是用生抽再加入焦糖色或其他色素制成的，颜色更深，呈棕褐色，味道相对较咸，鲜味不如生抽明显。老抽的主要用途是给食物上色，比如在红烧肉类、卤味或者调色时使用，让菜肴看起来更有食欲。\n \n不过，可能有些细节需要确认。比如，老抽是否真的添加了焦糖色？不同品牌的做法是否一致？另外，生抽和老抽在酿造工艺上的具体区别是什么？是否都经过了发酵过程，只是提取的方式不同？\n \n还有，用户可能想知道在实际烹饪中如何合理使用这两种酱油，以达到最佳的味道和颜色效果。例如，在炒青菜时，只需用少量生抽提鲜；而在做红烧肉时，则需要在翻炒肉块后加入老抽，使肉块均匀上色，之后再加入生抽和其他调料继续炖煮，这样既能保证肉质的鲜嫩，又能使菜肴的颜色看起来非常诱人。\n \n总结一下，生抽和老抽的主要区别在于颜色、味道和使用场景。生抽颜色浅、味道鲜，主要用于提鲜和调色；老抽颜色深、味道咸，主要用于给食物上色，使菜肴看起来更加诱人。在实际烹饪中，合理搭配使用这两种酱油，可以显著提升菜肴的口感和视觉效果。\n\n \n生抽与老抽是常见的酱油种类，它们在颜色、味道和使用场景上有显著区别：\n \n### **1. 颜色差异**\n- **生抽**：颜色较浅，呈红褐色或琥珀色。\n- **老抽**：颜色更深，呈棕褐色或黑褐色，类似焦糖色。\n \n### **2. 味道差异**\n- **生抽**：味道偏鲜，含较多氨基酸，常用于提鲜。\n- **老抽**：味道偏咸，鲜味较弱，主要作用是上色。\n \n### **3. 使用场景差异**\n- **生抽**：\n  - 炒菜、凉拌菜（如青菜、豆腐、凉拌鸡等）。\n  - 腌制食材（如腌黄瓜、泡菜等）。\n  - 蘸食（如白灼虾、蒸鱼等）。\n \n- **老抽**：\n  - 上色（如红烧类、卤味类、酱烧类等）。\n  - 调色（如制作汤底、酱料，或需要深色的菜肴）。\n  - 少量提味（虽然主要作用是上色，但少量使用可以增加菜肴的层次感）。\n \n### **4. 烹饪技巧**\n- **生抽**：\n  - 避免过量使用，以免掩盖食材本身的鲜味。\n  - 在凉拌菜中，可先将生抽与香油、蒜末等混合，再淋在食材上，这样味道更均匀。\n \n- **老抽**：\n  - 上色时，应早加入锅中，并翻炒均匀，使食材均匀上色。\n  - 避免在收汁阶段加入老抽，否则可能导致颜色过深，且不易均匀分布。\n  - 在制作红烧类菜肴时，可将老抽与生抽、糖、料酒等调味料提前调成汁，再倒入锅中与食材一同烧制，这样能更好地控制调味料的用量和火候，使菜肴的味道更加协调。\n \n### **总结**\n生抽与老抽的核心区别在于**颜色深浅**和**味道鲜咸**。生抽颜色浅、味道鲜，主要用于提鲜和轻上色；老抽颜色深、味道咸，主要用于重上色和轻微提味。在烹饪实践中，合理搭配使用这两种酱油，能够显著提升菜肴的口感层次和视觉吸引力。\n",[724],{"type":18,"tag":409,"props":725,"children":726},{"__ignoreMap":7},[727],{"type":24,"value":722},{"type":18,"tag":26,"props":729,"children":730},{},[731],{"type":18,"tag":78,"props":732,"children":733},{},[734],{"type":24,"value":735},"# 04",{"type":18,"tag":26,"props":737,"children":738},{},[739],{"type":18,"tag":78,"props":740,"children":741},{},[742],{"type":24,"value":743},"离线推理",{"type":18,"tag":26,"props":745,"children":746},{},[747,751],{"type":18,"tag":78,"props":748,"children":749},{},[750],{"type":24,"value":574},{"type":24,"value":752}," 在离线推理模式下，无需事先启动推理服务，每次执行推理脚本均会单独执行推理过程输出结果。",{"type":18,"tag":26,"props":754,"children":755},{},[756],{"type":18,"tag":78,"props":757,"children":758},{},[759],{"type":24,"value":98},{"type":18,"tag":26,"props":761,"children":762},{},[763],{"type":18,"tag":78,"props":764,"children":765},{},[766],{"type":24,"value":591},{"type":18,"tag":404,"props":768,"children":769},{"code":594},[770],{"type":18,"tag":409,"props":771,"children":772},{"__ignoreMap":7},[773],{"type":24,"value":594},{"type":18,"tag":26,"props":775,"children":776},{},[777],{"type":24,"value":189},{"type":18,"tag":26,"props":779,"children":780},{},[781],{"type":18,"tag":78,"props":782,"children":783},{},[784],{"type":24,"value":785},"执行以下离线推理python脚本：",{"type":18,"tag":404,"props":787,"children":789},{"code":788},"\nimport vllm_mindspore\nfrom vllm import LLM, SamplingParams\nfrom mindformers import AutoTokenizer\n \nif __name__ == \"__main__\":    \n    model='/home/teleAI/T1-35B' # 指定模型路径\n    tokenizer = AutoTokenizer.from_pretrained(model, trust_remote_code=True)\n    sampling_params = SamplingParams(temperature=0.6, repetition_penalty=1.05, max_tokens=8192)\n    llm = LLM(model=model, trust_remote_code=True, tensor_parallel_size=4)\n \n    prompt = \"生抽与老抽的区别？\"\n    messages = [{\"role\": \"user\", \"content\": prompt}]\n    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n \n    outputs = llm.generate([text], sampling_params)\n    for output in outputs:\n        prompt = output.prompt\n        generated_text = output.outputs[0].text\n        print(f\"Prompt: {prompt!r},\nGenerated text: {generated_text!r}\")\n",[790],{"type":18,"tag":409,"props":791,"children":792},{"__ignoreMap":7},[793],{"type":24,"value":788},{"type":18,"tag":26,"props":795,"children":796},{},[797],{"type":18,"tag":78,"props":798,"children":799},{},[800],{"type":24,"value":801},"脚本说明：",{"type":18,"tag":152,"props":803,"children":804},{},[805,810,815,831],{"type":18,"tag":156,"props":806,"children":807},{},[808],{"type":24,"value":809},"T1 系列模型在 chat template 中加入了\\n符号以确保推理时能够生成 reason 过程。推理脚本会自动在推理起始拼接\\n符号，此时输出结果会缺少开头的\\n符号。",{"type":18,"tag":156,"props":811,"children":812},{},[813],{"type":24,"value":814},"model 需要配置为实际的模型路径。",{"type":18,"tag":156,"props":816,"children":817},{},[818,820,824,826,829],{"type":24,"value":819},"sampling_params 推理参数配置：",{"type":18,"tag":821,"props":822,"children":823},"br",{},[],{"type":24,"value":825},"1、在推理数学、代码任务时，建议使用repetition_penalty=1.0, temperature=0.6, top_p=0.95的推理设置。",{"type":18,"tag":821,"props":827,"children":828},{},[],{"type":24,"value":830},"2、在推理通用任务时，建议使用repetition_penalty=1.05, temperature=0.6, top_p=0.95的推理设置，可以有效减少重复生成现象。",{"type":18,"tag":156,"props":832,"children":833},{},[834],{"type":24,"value":835},"prompt 可修改为其他推理问题。",{"type":18,"tag":26,"props":837,"children":838},{},[839],{"type":18,"tag":78,"props":840,"children":841},{},[842],{"type":24,"value":501},{"type":18,"tag":26,"props":844,"children":845},{},[846],{"type":18,"tag":78,"props":847,"children":848},{},[849],{"type":24,"value":850},"推理结果",{"type":18,"tag":404,"props":852,"children":854},{"code":853},"嗯，用户问生抽和老抽的区别。首先，我得确认自己对这两个调味品的了解是否正确。\n\n生抽，我记得主要是用于炒菜和凉拌的。它的颜色较浅，呈红褐色，味道比较咸，但带有一定的鲜味，因为通常含有谷氨酸钠（味精）。\n\n然后是老抽，主要用于给食物上色，比如红烧、卤味等。老抽的颜色更深，呈深红或黑红色。味道方面，老抽比生抽更咸，且鲜味相对较少，因为有些老抽可能不含味精，或者含量较低。\n\n接下来，用户可能还关心的是，这两种酱油在烹饪时的具体应用，以及它们对菜肴风味和色泽的影响。此外，可能还需要提到一些关于酱油制作的传统工艺，比如发酵时间、原料配比等，这些都会影响生抽和老抽的最终风味和质地。\n\n不过，在回答用户的问题时，需要保持回答的简洁性和针对性，避免过于冗长或偏离主题。因此，在总结生抽和老抽的区别时，应该重点突出它们在颜色、用途、风味等方面的不同，同时也可以简要提及它们在烹饪中的具体应用场景，以及它们对菜肴整体风味和视觉效果的影响。\n\n最后，在确保回答准确无误的基础上，可以用一种较为亲切和自然的方式将这些信息呈现给用户，让用户能够轻松理解并记住生抽和老抽之间的主要区别。\n\n\n生抽与老抽都是常见的酱油品种，但它们在颜色、用途、风味等方面有显著区别：\n---\n\n### **1. 颜色差异**\n- **生抽**：颜色较浅，呈红褐色或琥珀色。  \n- **老抽**：颜色深，呈深红、黑红或接近黑色。  \n\n---\n\n### **2. 用途差异**\n- **生抽**：主要用于日常炒菜、凉拌、蘸食等，提鲜增香。  \n- **老抽**：主要用于给菜肴上色，例如红烧、卤味、酱烧等，使菜品色泽更诱人。  \n\n---\n\n### **3. 风味差异**\n- **生抽**：味道较咸，但带有明显的鲜味（因含谷氨酸钠，即味精）。适合直接用于调味。  \n- **老抽**：味道更咸，鲜味相对较弱（部分老抽可能不含味精）。由于主要用于上色，因此在调味时通常不会直接使用老抽，而是在出锅前少量淋入以增色。  \n\n---\n\n### **4. 存储方式**\n- 两种酱油均需存放在阴凉干燥处，避免高温和阳光直射，以免加速变质或损失风味。  \n\n---\n\n### **总结对比表**\n| 特性       | 生抽                     | 老抽                     |\n|--------------|----------------------------|----------------------------|\n| **颜色**   | 浅红褐色                 | 深红/黑红色            |\n| **用途**     | 炒菜、凉拌、蘸食      | 上色（红烧、卤味等） |\n| **风味**     | 咸鲜味，含味精           | 咸味为主，鲜味较弱       |\n| **存储**     | 阴凉干燥处               | 同上                       |\n \n通过以上对比，可以清晰地理解生抽与老抽在烹饪中的不同角色和用途。\n",[855],{"type":18,"tag":409,"props":856,"children":857},{"__ignoreMap":7},[858],{"type":24,"value":853},{"type":18,"tag":26,"props":860,"children":861},{},[862],{"type":24,"value":863},"昇思MindSpore AI框架将持续支持相关主流模型演进，并根据情况向全体开发者提供镜像与支持。",{"title":7,"searchDepth":865,"depth":865,"links":866},4,[867,868,869],{"id":245,"depth":865,"text":245},{"id":261,"depth":865,"text":261},{"id":282,"depth":865,"text":282},"markdown","content:technology-blogs:zh:3723.md","content","technology-blogs/zh/3723.md","technology-blogs/zh/3723","md",1776506134111]