Jagorar Mai Amfani NeMo Framework NVIDIA

NVIDIA NeMo Framework User Guide

NVIDIA-logo

NVIDIA NeMo Framework

NVIDIA-NeMo-Framework-samfurin

Ƙayyadaddun bayanai

  • Sunan samfur: NVIDIA NeMo Framework
  • Dandali da abin ya shafa: Windows, Linux, macOS
  • Sigar da abin ya shafa: All versions kafin 24
  • Lalacewar Tsaro: CVE-2025-23360
  • Makin Ƙididdigar Haɗari: 7.1 (CVSS v3.1)

Umarnin Amfani da samfur

Shigar Sabunta Tsaro:
Don kare tsarin ku, bi waɗannan matakan:

  1. Zazzage sabon saki daga NeMo-Framework-Launcher Releases page akan GitHub.
  2. Jeka Tsaron Samfur na NVIDIA don ƙarin bayani.

Cikakkun Sabunta Tsaro:
Sabuntawar tsaro yana magance rashin lahani a cikin Tsarin NVIDIA NeMo wanda zai iya haifar da aiwatar da code da bayanan t.ampirin.

Inganta Software:
Idan kana amfani da sakin reshe na baya, ana ba da shawarar haɓaka zuwa sabon sakin reshe don magance matsalar tsaro.

Ƙarsheview

NVIDIA NeMo Tsarin tsari ne mai daidaitawa kuma tushen tushen tushen tushen AI wanda aka gina don masu bincike da masu haɓakawa waɗanda ke aiki akan. Manyan Samfuran Harshe, Multimodal, da Magana AI (misali Gane Magana ta atomatik kuma Rubutu-zuwa-Magana). Yana ba masu amfani damar ƙirƙira, keɓancewa, da tura sabbin samfuran AI na ƙirƙira ta hanyar haɓaka lambar data kasance da wuraren binciken samfurin da aka riga aka horar.

Saita UmarninShigar NeMo Framework

Manyan Samfuran Harshe da Model Multimodal
Tsarin NeMo yana ba da tallafi na ƙarshen-zuwa-ƙarshe don haɓaka Samfuran Harshe Manyan (LLMs) da Multimodal Model (MMs). Yana ba da sassaucin da za a yi amfani da shi a kan-gidaje, a cikin cibiyar bayanai, ko tare da mai samar da girgije da kuka fi so. Hakanan yana goyan bayan aiwatarwa akan wuraren kunna SLURM ko Kubernetes.

_images/nemo-llm-mm-stack.png

Ƙididdigar Bayanai
NeMo Curator [1] Laburaren Python ne wanda ya haɗa da rukunin kayayyaki don hakar bayanai da samar da bayanan roba. Suna iya daidaitawa kuma an inganta su don GPUs, yana sa su dace don sarrafa bayanan harshe na halitta don horarwa ko daidaita LLMs. Tare da NeMo Curator, zaku iya fitar da ingantaccen rubutu mai inganci daga ɗanyen mai yawa web tushen bayanai.

Horowa da Keɓancewa

Tsarin NeMo yana ba da kayan aikin don ingantaccen horarwa da keɓancewa na LLMs da Multimodal model. Ya haɗa da saitunan tsoho don saitin gungu, zazzage bayanai, da ƙirar hyperparameters, waɗanda za'a iya daidaita su don horar da sabbin bayanai da ƙira. Baya ga horarwa na farko, NeMo yana goyan bayan duka Dabaru Mai Kyau mai Kyau (SFT) da Dabarun Ingantaccen Gyaran Tunani (PEFT) kamar LoRA, Ptuning, da ƙari.

Akwai zaɓuɓɓuka biyu don ƙaddamar da horo a cikin NeMo - ta amfani da NeMo 2.0 API dubawa ko tare da NeMo Run.

  • Tare da NeMo Run (An shawarta): NeMo Run yana ba da hanyar sadarwa don daidaita tsari, aiwatarwa da gudanar da gwaje-gwaje a cikin mahallin ƙididdiga daban-daban. Wannan ya haɗa da ƙaddamar da ayyuka a kan aikinku na gida ko kan manyan gungu - duka SLURM an kunna ko Kubernetes a cikin yanayin girgije.
    • Horowar farko & PEFT Quickstart tare da NeMo Run
  • Amfani da NeMo 2.0 API: Wannan hanyar tana aiki da kyau tare da saiti mai sauƙi wanda ya ƙunshi ƙananan ƙira, ko kuma idan kuna sha'awar rubuta naku mai ɗaukar bayanan al'ada, madaukai horo, ko canza yadudduka samfurin. Yana ba ku ƙarin sassauci da iko akan daidaitawa, kuma yana sauƙaƙa don tsawaitawa da tsara tsarin daidaitawa ta hanyar tsari.
    • TraIning Quickstart tare da NeMo 2.0 API
    • Hijira daga NeMo 1.0 zuwa NeMo 2.0 API

Daidaitawa

  • NeMo-Aligner [1] Kit ɗin kayan aiki ne mai ƙima don daidaita ƙirar ƙira mai inganci. Kayan aikin yana da goyan baya ga ƙirar ƙirar ƙirar ƙirar zamani kamar SteerLM, DPO, Ƙarfafa Koyo daga Feedback na ɗan adam (RLHF), da ƙari mai yawa. Waɗannan algorithms suna ba masu amfani damar daidaita ƙirar harshe don zama mafi aminci, mara lahani, da taimako.
  • Duk wuraren bincike na NeMo-Aligner sun dace da tsarin halittu na NeMo, suna ba da damar ƙarin gyare-gyare da ƙaddamar da ƙaddamarwa.

Gudun aikin mataki-mataki na duk matakai uku na RLHF akan ƙaramin ƙirar GPT-2B:

  • Farashin SFT
  • Horon samfurin kyauta
  • Horon PPO

Bugu da kari, muna nuna goyan baya ga wasu hanyoyin daidaitawa daban-daban:

  • DPO: algorithm daidaita nauyi mai nauyi idan aka kwatanta da RLHF tare da aikin asara mai sauƙi.
  • Wasa-kai Kyakkyawan Tunani (SPIN)
  • SteerLM: wata dabara da ta dogara akan sharadi-SFT, tare da fitarwa mai iya jurewa.

Duba takaddun don ƙarin bayani: Takardun Daidaitawa

Multimodal Model

  • Tsarin tsari na Nemo yana samar da ingantaccen software don horarwa da kuma ƙirar multory na zamani, tushe na harshe, kuma bayan ƙarni na rubutu ta amfani da filayen 2D.
  • An ƙera kowane nau'i ne don biyan takamaiman buƙatu da ci gaba a fagen, yin amfani da ƙirar ƙira don ɗaukar nau'ikan nau'ikan bayanai da yawa, gami da rubutu, hotuna, da ƙirar 3D.

Lura
Muna ƙaura tallafi don ƙirar multimodal daga NeMo 1.0 zuwa NeMo 2.0. Idan kuna son bincika wannan yanki a halin yanzu, da fatan za a koma zuwa takaddun don sakin NeMo 24.07 (a baya).

Ƙaddamarwa da Ƙaddamarwa
Tsarin NeMo yana ba da hanyoyi daban-daban don ƙaddamar da LLM, yana ba da yanayi daban-daban na turawa da buƙatun aiki.

Yi amfani da NVIDIA NIM

  • Tsarin NeMo ba tare da matsala ba yana haɗawa tare da kayan aikin tura samfurin matakin kasuwanci ta hanyar NVIDIA NIM. NVIDIA TensorRT-LLM ce ke ba da wannan haɗin kai, yana tabbatar da ingantacciyar ƙima da ƙima.
  • Don ƙarin bayani kan NIM, ziyarci NVIDIA website.

Yi aiki tare da TensorRT-LLM ko vLLM

  • Tsarin NeMo yana ba da rubutun rubutu da APIs don fitarwa samfura zuwa ingantattun ɗakunan karatu guda biyu, TensorRT-LLM da vLLM, kuma don tura ƙirar da aka fitar tare da NVIDIA Triton Inference Server.
  • Don yanayin yanayin da ke buƙatar ingantaccen aiki, ƙirar NeMo na iya yin amfani da TensorRT-LLM, ɗakin karatu na musamman don haɓakawa da haɓaka ƙimar LLM akan NVIDIA GPUs. Wannan tsari ya ƙunshi canza ƙirar NeMo zuwa tsarin da ya dace da TensorRT-LLM ta amfani da tsarin nemo.export.
    • LLM ya ƙareview
    • Sanya Manyan Samfuran Harshe NeMo tare da NIM
    • Ƙarfafa Samfuran Manyan Harshe NeMo tare da TensorRT-LLM
    • Sanya Manyan Samfuran Harshe na NeMo tare da vLLM

Samfuran Tallafi

Manyan Samfuran Harshe

Manyan Samfuran Harshe
Manyan Samfuran Harshe Pretraining & SFT PEFT Daidaitawa Haɗuwar Horon FP8 TRT/TRTLLM Juya Zuwa & Daga Fuskar Runguma Kimantawa
Llama3 8B/70B, Llama3.1 405B Ee Ee x Ee (an tabbatar da wani bangare) Ee Duka Ee
Mixtral 8x7B/8x22B Ee Ee x Ee (ba a tantance ba) Ee Duka Ee
Nemotron 3B Ee x x Ee (ba a tantance ba) x Duka Ee
Nemotron 4B Ee x x Ee (ba a tantance ba) x Duka Ee
Baichuan 2B Ee Ee x Ee (ba a tantance ba) x Duka Ee
Saukewa: GLM3B Ee Ee x Ee (ba a tantance ba) x Duka Ee
Gemma 2B/7B Ee Ee x Ee (ba a tantance ba) Ee Duka Ee
Gemma2 2B/9B/27B Ee Ee x Ee (ba a tantance ba) x Duka Ee
Mamba2 130M/370M/780M/1.3B/2.7B/8B/ Hybrid-8B Ee Ee x Ee (ba a tantance ba) x x Ee
Phi3 mini 4k x Ee x Ee (ba a tantance ba) x x x
Qwen2 0.5B/1.5B/7B/72B Ee Ee x Ee (ba a tantance ba) Ee Duka Ee
StarCoder 15B Ee Ee x Ee (ba a tantance ba) Ee Duka Ee
StarCoder2 3B/7B/15B Ee Ee x Ee (ba a tantance ba) Ee Duka Ee
BERT 110M/340M Ee Ee x Ee (ba a tantance ba) x Duka x
T5 220M/3B/11B Ee Ee x x x x x

 

Samfuran Harshen Hangen nesa

Samfuran Harshen Hangen nesa
Samfuran Harshen Hangen nesa Pretraining & SFT PEFT Daidaitawa Haɗuwar Horon FP8 TRT/TRTLLM Juya Zuwa & Daga Fuskar Runguma Kimantawa
NeVA (LLaVA 1.5) Ee Ee x Ee (ba a tantance ba) x Daga x
Llama 3.2 Vision 11B/90B Ee Ee x Ee (ba a tantance ba) x Daga x
LLaVA na gaba (LLaVA 1.6) Ee Ee x Ee (ba a tantance ba) x Daga x

 

Abubuwan Haɗawa

Abubuwan Haɗawa
Haɗa Samfuran Harshe Pretraining & SFT PEFT Daidaitawa Haɗuwar Horon FP8 TRT/TRTLLM Juya Zuwa & Daga Fuskar Runguma Kimantawa
SBERT 340M Ee x x Ee (ba a tantance ba) x Duka x
Llama 3.2 Shigar 1B Ee x x Ee (ba a tantance ba) x Duka x

 

Samfuran Gidauniyar Duniya

Samfuran Gidauniyar Duniya
Samfuran Gidauniyar Duniya Bayan horo Gaggauta Fitowa
Cosmos-1.0-Yaduwa-Text2World-7B Ee Ee
Cosmos-1.0-Yaduwa-Text2World-14B Ee Ee
Cosmos-1.0-Yaduwa-Video2World-7B Ana zuwa Nan ba da jimawa ba Ana zuwa Nan ba da jimawa ba
Cosmos-1.0-Yaduwa-Video2World-14B Ana zuwa Nan ba da jimawa ba Ana zuwa Nan ba da jimawa ba
Cosmos-1.0-Autoregressive-4B Ee Ee
Cosmos-1.0-Autoregressive-Video2World-5B Ana zuwa Nan ba da jimawa ba Ana zuwa Nan ba da jimawa ba
Cosmos-1.0-Autoregressive-12B Ee Ee
Cosmos-1.0-Autoregressive-Video2World-13B Ana zuwa Nan ba da jimawa ba Ana zuwa Nan ba da jimawa ba

Lura
NeMo kuma yana goyan bayan pretraining don yaduwa da gine-ginen autoregressive text2world kafuwar model.

Magana AI

Haɓaka samfuran AI na tattaunawa wani tsari ne mai rikitarwa wanda ya haɗa da ma'ana, gini, da ƙirar horarwa a cikin takamaiman yankuna. Wannan tsari yawanci yana buƙatar maimaitawa da yawa don isa babban matakin daidaito. Yakan haɗa da maimaitawa da yawa don cimma daidaitattun daidaito, daidaitawa akan ayyuka daban-daban da takamaiman bayanai na yanki, tabbatar da aikin horo, da shirya samfura don ƙaddamar da ƙaddamarwa.

_images/nemo-magana-ai.png

Tsarin NeMo yana ba da tallafi don horarwa da gyare-gyaren ƙirar Magana AI. Wannan ya haɗa da ayyuka kamar Gane Magana ta atomatik (ASR) da haɗin rubutu-zuwa-Magana (TTS). Yana ba da sauƙi mai sauƙi zuwa ƙaddamar da matakin samar da kasuwanci tare da NVIDIA Riva. Don taimakawa masu haɓakawa da masu bincike, Tsarin NeMo ya haɗa da wuraren bincike na zamani da aka riga aka horar da su, kayan aiki don sarrafa bayanan maganganun da za a iya maimaitawa, da fasali don bincike mai mu'amala da nazarin bayanan bayanan magana. Abubuwan da ke cikin Tsarin NeMo don Magana AI sune kamar haka:

Horowa da Keɓancewa
Tsarin NeMo ya ƙunshi duk abin da ake buƙata don horarwa da tsara ƙirar magana (ASRRarraba MaganaGane Kakakin MajalisaDiarization Speaker, kuma TTS) ta hanyar da za a iya haifuwa.

Samfuran da aka riga aka horar da SOTA

  • Tsarin NeMo yana ba da girke-girke na zamani da wuraren bincike da aka riga aka horar da su ASR kuma TTS model, kazalika da umarnin kan yadda za a loda su.
  • Kayan Aikin Magana
  • Tsarin NeMo yana ba da saitin kayan aiki masu amfani don haɓaka samfuran ASR da TTS, gami da:
    • NeMo Forced Aligner (NFA) don samar da alamar-, kalma- da lokutan matakin-ɓangareamps na magana a cikin sauti ta amfani da ƙirar Gane Magana ta atomatik na tushen NeMo na CTC.
    • Mai sarrafa Bayanan Magana (SDP), kayan aiki don sauƙaƙe sarrafa bayanan magana. Yana ba ku damar wakiltar ayyukan sarrafa bayanai a cikin saiti file, rage girman lambar tukunyar jirgi da ba da damar sake haifuwa da rabo.
    • Mai Binciken Magana (SDE), tushen Dash web aikace-aikacen bincike na mu'amala da nazarin bayanan bayanan magana.
    • Kayan aikin ƙirƙirar bayanai wanda ke ba da ayyuka don daidaita doguwar sauti files tare da madaidaitan kwafin kuma raba su zuwa guntun guntu waɗanda suka dace da horarwar ƙirar ƙira ta atomatik (ASR).
    • Kayan Aikin Kwatance don Model ASR don kwatanta tsinkaya na nau'ikan ASR daban-daban a daidaitattun kalmomi da matakin furci.
    • ASR Evaluator don kimanta aikin samfuran ASR da sauran fasalulluka kamar Gane Ayyukan Murya.
    • Kayan aikin Daidaita Rubutu don canza rubutu daga rubutaccen tsari zuwa sigar magana da akasin haka (misali "31st" da "talatin farko").
  • Hanyar Tuba
  • Samfuran NeMo waɗanda aka horar da su ko keɓance su ta amfani da Tsarin NeMo za a iya inganta su kuma a tura su tare da NVIDIA Riva. Riva yana ba da kwantena da taswirar Helm da aka tsara musamman don sarrafa matakan aiwatar da maɓallin turawa.

Sauran Albarkatun

GitHub Repos
  • NeMo: Babban wurin ajiya na Tsarin NeMo
  • NeMo-Gudu: Kayan aiki don daidaitawa, ƙaddamarwa da sarrafa gwaje-gwajen koyon injin ku.
  • NeMo-Aligner: Kayan aiki mai ƙididdigewa don ingantaccen jeri na ƙira
  • NeMo-Curator: Abubuwan da za a iya ƙididdigewa kafin aiwatarwa da kayan aikin kayan aiki don LLMs
Samun Taimako
Yi hulɗa tare da al'ummar NeMo, yi tambayoyi, sami tallafi, ko bayar da rahoton kwari.
  • Tattaunawar NeMo
  • Abubuwan NeMo

Harsunan Shirye-shiryen da Tsarin Tsarin

  • Python: Babban hanyar sadarwa don amfani da Tsarin NeMo
  • Pytorch: An gina tsarin NeMo a saman PyTorch

Lasisi

  • NeMo Github repo yana da lasisi a ƙarƙashin lasisin Apache 2.0
  • NeMo Framework yana da lasisi a ƙarƙashin NVIDIA AI PRODUCT GREEMENT. Ta ja da amfani da akwati, kun karɓi sharuɗɗa da sharuɗɗan wannan lasisi.
  • Kwandon Tsarin NeMo ya ƙunshi kayan Llama waɗanda ke ƙarƙashin Yarjejeniyar Lasisi na Al'umma Meta Llama3.

Bayanan kafa
A halin yanzu, NeMo Curator da NeMo Aligner goyon bayan Multimodal model aiki ne na ci gaba kuma za a samu nan ba da jimawa ba.

FAQ

Tambaya: Ta yaya zan iya bincika idan rauni ya shafi tsarina?
A: Kuna iya bincika idan tsarinku ya shafi ta hanyar tabbatar da sigar NVIDIA NeMo Framework da aka shigar. Idan yana ƙasa da sigar 24, tsarin ku na iya zama mai rauni.

Tambaya: Wanene ya ba da rahoton batun tsaro CVE-2025-23360?
A: Ko Peles - JFrog Tsaro ya ruwaito batun tsaro. NVIDIA ta amince da gudunmawarsu.

Tambaya: Ta yaya zan iya karɓar sanarwar sanarwar tsaro nan gaba?
A: Ziyarci shafin Tsaron Samfur na NVIDIA don biyan kuɗi zuwa sanarwar sanarwar tsaro kuma ku kasance da masaniya game da sabunta tsaro na samfur.

Takardu / Albarkatu

PDF thumbnailNeMo Tsarin
User Guide · NeMo Framework, NeMo, Framework

Magana

Yi Tambaya

Use this section to ask about setup, compatibility, troubleshooting, or anything missing from this manual.

Yi Tambaya

Ask about setup, compatibility, troubleshooting, or anything missing from this manual. Name and email are optional.