NVIDIA NeMo Framework

Fa'amatalaga

• Igoa Oloa: NVIDIA NeMo Framework
• Fa'aa'afia Fa'atonu: Pupuni, Linux, macOS
• Ituaiga A'afia: Fa'aliliuga uma a'o lumana'i le 24
• Fa'aletonu le saogalemu: CVE-2025-23360
• Fuafuaga Fa'avae Iloiloga Tulaga lamatia: 7.1 (CVSS v3.1)

Fa'atonuga o le Fa'aaogaina o Mea

Fa'atonuga Fa'afouga Saogalemu:
Ina ia puipuia lau polokalama, mulimuli i laasaga nei:

1. La'uina le fa'asalalauga lata mai mai le NeMo-Framework-Launcher Releases page ile GitHub.
2. Alu ile NVIDIA Product Security mo nisi fa'amatalaga.

Fa'amatalaga Fa'afouga Saogalemu:
O le faʻafouina o le puipuiga e faʻaalia ai se faʻafitauli i le NVIDIA NeMo Framework e mafai ona taʻitaʻia ai le faʻatinoina o tulafono ma faʻamaumauga t.ampsese.

Faʻaleleia polokalama:
Afai o loʻo e faʻaaogaina se faʻasalalauga muamua o le lala, e fautuaina e faʻaleleia i le faʻasalalauga lata mai o le lala e foia ai le faafitauli tau puipuiga.

Ua umaview

NVIDIA NeMo Framework ose fa'avaeina fa'atupuina AI fa'avae ma fa'aoso fa'atupuina mo tagata su'esu'e ma atina'e o lo'o galulue i luga. Fa'ata'ita'iga Gagana Tele, Multimodal, ma Lauga AI (eg Otometi le Fa'ailoaina o Tautala ma Tusi-i-Lauga). E mafai ai e tagata fa'aoga ona fa'atino lelei, fa'avasega, ma fa'apipi'i fa'ata'ita'iga fa'atupu AI fou e ala i le fa'aogaina o fa'ailoga o lo'o iai ma fa'ata'ita'iga fa'ata'ita'i a'oa'oga.

Seti Faatonuga: Faʻapipiʻi le NeMo Framework

Fa'ata'ita'iga Gagana Tele ma Fa'ata'ita'iga Multimodal
NeMo Framework e tu'uina atu le lagolago pito i'u mo le atina'eina o Fa'ata'ita'iga Gagana Tele (LLMs) ma Fa'ata'ita'iga Multimodal (MMs). E maua ai le fetuutuuna'i e fa'aoga i luga ole fale, ile nofoaga autu o fa'amaumauga, po'o le tu'uina atu o ao e te mana'omia. E lagolagoina fo'i le fa'atinoina i SLURM po'o Kubernetes si'osi'omaga mafai.

Fa'amaumauga Fa'amaumauga
NeMo Pule [1] o se faletusi Python e aofia ai se seti o modules mo faʻamatalaga faʻamaumauga ma faʻamaumauga faʻapipiʻi. E mafai ona fa'avasegaina ma fa'amalieina mo GPU, e fa'amalieina mo le fa'avasegaina o fa'amatalaga o gagana fa'ale-natura e toleni ai pe fa'alelei lelei LLMs. Faatasi ai ma le NeMo Curator, e mafai ona e suʻeina lelei tusitusiga maualuga mai le tele o mata web puna'oa fa'amaumauga.

A'oa'oga ma Fa'asinomaga

NeMo Framework e tu'uina atu meafaigaluega mo a'oa'oga lelei ma fa'avasegaina o LLMs ma Multimodal faʻataʻitaʻiga. E aofia ai fa'atonuga fa'aletonu mo le fa'atulagaina o fa'aputuga fa'aputu, la'uina o fa'amaumauga, ma fa'ata'ita'iga hyperparameters, lea e mafai ona fetuuna'i e a'oa'oina i fa'amaumauga fou ma fa'ata'ita'iga. I le faaopoopo atu i le muaʻi aʻoaʻoga, o loʻo lagolagoina e NeMo ia faʻataʻitaʻiga Suavaia Fine-Tuning (SFT) ma Parameter Efficient Fine-Tuning (PEFT) e pei o LoRA, Ptuning, ma isi mea.

E lua filifiliga o loʻo avanoa e faʻalauiloa aʻoaʻoga i NeMo - faʻaaoga le NeMo 2.0 API faʻaoga poʻo le NeMo Run.

• Faatasi ai ma le NeMo Run (Fautuaina): NeMo Run e maua ai se atina'e e fa'afaigofie ai le fa'atulagaina, fa'atinoga ma le fa'afoega o fa'ata'ita'iga i si'osi'omaga tau fa'atatau. E aofia ai le tatalaina o galuega i luga o lau fale faigaluega i le lotoifale poʻo luga o fuifui tetele - uma SLURM faʻatagaina poʻo Kubernetes i totonu ole siosiomaga ao.
  • A'o le a'oa'oga & Amata vave PEFT ma le NeMo Run
• Fa'aaogā le NeMo 2.0 API: O lenei metotia e aoga lelei i se seti faigofie e aofia ai faʻataʻitaʻiga laiti, pe afai e te fiafia e tusi lau lava faʻamaumauga masani, faʻaoga faʻaoga, poʻo le suia o laulau faʻataʻitaʻiga. E tu'uina atu ia te oe le sili atu le fetuutuunai ma le pulea o fa'atonuga, ma fa'afaigofie ona fa'alautele ma fa'avasega fa'apolokalame.
  • Training Quickstart ma NeMo 2.0 API
  • Femalaga mai NeMo 1.0 i le NeMo 2.0 API

Fa'atonuga

• NeMo-Aligner [1] ose mea faigaluega e mafai ona fuaina mo le fa'aogaina lelei o fa'ata'ita'iga. O lo'o i ai i le pusa meafaigaluega le lagolago mo fa'ata'ita'iga fa'ata'ita'iga fa'ata'ita'i fa'ata'ita'iga algorithms e pei ole SteerLM, DPO, Fa'amalosi A'oa'oga mai Manatu Fa'afofoga a Tagata (RLHF), ma le tele o isi mea. O nei algorithms e mafai ai e tagata faʻaoga ona faʻaogaina faʻataʻitaʻiga gagana ina ia sili atu le saogalemu, le afaina, ma fesoasoani.
• O siaki uma a NeMo-Aligner e fesoʻotaʻi faʻatasi ma le NeMo ecosystem, faʻatagaina mo le faʻaogaina atili ma le faʻatulagaina o faʻamatalaga.

Laasaga taʻitasi laʻasaga galue o vaega uma e tolu o le RLHF i luga o se tamaʻi GPT-2B faʻataʻitaʻiga:

• SFT aoaoga
• Tauia faʻataʻitaʻiga aʻoaʻoga
• a'oa'oga PPO

E le gata i lea, matou te faʻaalia le lagolago mo isi auala faʻaogaina fou:

• DPO: o se algorithm alignment mama pe a faatusatusa i le RLHF ma se galuega leiloa faigofie.
• Ta'aloga a le tagata lava ia Fa'alogo lelei (SPIN)
• SteerLM: o se metotia faʻavae i luga ole conditioned-SFT, faʻatasi ai ma mea e mafai ona faʻaogaina.

Siaki le pepa mo nisi faʻamatalaga: Fa'asologa o Fa'amaumauga

Multimodal Models

• NeMo Framework o lo'o tu'uina atu polokalame fa'apitoa e a'oa'o ma fa'apipi'i fa'ata'ita'iga fa'ata'ita'i fa'aonaponei i le tele o vaega: Fa'ata'ita'iga Gagana Tele, Fa'avae Va'aiga-Language, Fa'ata'ita'iga Fa'amatalaga-i-Ata, ma tua atu o le 2D Generation e fa'aaoga ai Neural Radiance Fields (NeRF).
• O vaega taʻitasi ua mamanuina e faʻafetaui ai manaʻoga faʻapitoa ma le alualu i luma i le fanua, faʻaogaina faʻataʻitaʻiga faʻataʻitaʻiga e faʻatautaia ai le tele o ituaiga faʻamaumauga, e aofia ai tusitusiga, ata, ma faʻataʻitaʻiga 3D.

Fa'atino ma Fa'ai'uga
NeMo Framework o lo'o tu'uina atu auala eseese mo le LLM fa'ai'uga, fa'afeiloa'i i fa'ata'ita'iga fa'atino eseese ma mana'oga fa'atinoga.

Fa'aoga ile NVIDIA NIM

• NeMo Framework o lo'o tu'ufa'atasia fa'atasi ma mea faigaluega fa'ata'ita'iga fa'ata'ita'iga ile atina'e ile NVIDIA NIM. O lenei tu'ufa'atasiga o lo'o fa'amalosia e le NVIDIA TensorRT-LLM, fa'amautinoaina le fa'amalieina ma le fa'aogaina o fa'amatalaga.
• Mo nisi faʻamatalaga ile NIM, asiasi ile NVIDIA webnofoaga.

Fa'aoga ile TensorRT-LLM po'o le vLLM

• NeMo Framework e ofoina atu tusitusiga ma API e auina atu ai faʻataʻitaʻiga i faletusi e lua e sili ona lelei, TensorRT-LLM ma vLLM, ma faʻapipiʻi le faʻataʻitaʻiga auina atu i fafo ma le NVIDIA Triton Inference Server.
• Mo faʻataʻitaʻiga e manaʻomia ai le faʻaogaina lelei, o NeMo faʻataʻitaʻiga e mafai ona faʻaogaina TensorRT-LLM, o se faletusi faʻapitoa mo le faʻavaveina ma le faʻamalieina o le LLM i luga ole NVIDIA GPU. O lenei faiga e aofia ai le fa'aliliuina o fa'ata'ita'iga NeMo i se fa'asologa e fetaui ma le TensorRT-LLM fa'aaoga le nemo.export module.
  • LLM Fa'atinoinaview
  • Fa'ata'ita'i Fa'ata'ita'iga Gagana Tele NeMo ma le NIM
  • Fa'ata'ita'i Fa'ata'ita'iga Gagana Tele NeMo ma TensorRT-LLM
  • Fa'ata'ita'i Fa'ata'ita'iga Gagana Tele a NeMo ma vLLM

Fa'ata'ita'iga Lagolago

Fa'ata'ita'iga Gagana Tele

Fa'ata'ita'iga Gagana Tele

Fa'ata'ita'iga Gagana Tele	mua'i a'oa'oga & SFT	PEFT	Fa'atonuga	FP8 Aoaoga Fa'atasi	TRT/TRTLLM	Su'e I & Mai Usuli Mata	Iloiloga
Llama3 8B/70B, Llama3.1 405B	Ioe	Ioe	x	Ioe (fa'amaonia vaega)	Ioe	E lua	Ioe
Fa'afefiloi 8x7B/8x22B	Ioe	Ioe	x	Ioe (le fa'amaonia)	Ioe	E lua	Ioe
Nemotron 3 8B	Ioe	x	x	Ioe (le fa'amaonia)	x	E lua	Ioe
Nemotron 4 340B	Ioe	x	x	Ioe (le fa'amaonia)	x	E lua	Ioe
Baichuan2 7B	Ioe	Ioe	x	Ioe (le fa'amaonia)	x	E lua	Ioe
TalatalaGLM3 6B	Ioe	Ioe	x	Ioe (le fa'amaonia)	x	E lua	Ioe
Gemma 2B/7B	Ioe	Ioe	x	Ioe (le fa'amaonia)	Ioe	E lua	Ioe
Gemma2 2B/9B/27B	Ioe	Ioe	x	Ioe (le fa'amaonia)	x	E lua	Ioe
Mamba2 130M/370M/780M/1.3B/2.7B/8B/ Hybrid-8B	Ioe	Ioe	x	Ioe (le fa'amaonia)	x	x	Ioe
Phi3 mini 4k	x	Ioe	x	Ioe (le fa'amaonia)	x	x	x
Qwen2 0.5B/1.5B/7B/72B	Ioe	Ioe	x	Ioe (le fa'amaonia)	Ioe	E lua	Ioe
FetuCoder 15B	Ioe	Ioe	x	Ioe (le fa'amaonia)	Ioe	E lua	Ioe
StarCoder2 3B/7B/15B	Ioe	Ioe	x	Ioe (le fa'amaonia)	Ioe	E lua	Ioe
BERT 110M/340M	Ioe	Ioe	x	Ioe (le fa'amaonia)	x	E lua	x
T5 220M/3B/11B	Ioe	Ioe	x	x	x	x	x

 

Fa'ata'ita'iga Gagana Va'aiga

 

Fa'apipi'i Fa'ata'ita'iga

 

Fa'ata'ita'iga Fa'avae a le Lalolagi

Manatua
E lagolagoina foi e NeMo le muaʻi aʻoaʻoina mo le faʻasalalauina ma le autoregressive architectures text2world fa'ata'ita'iga fa'avae.

Lauga AI

O le atina'eina o fa'ata'ita'iga fa'ata'ita'iga AI ose faiga fa'alavelave e aofia ai le fa'amalamalamaina, fau, ma fa'aa'oa'oga fa'ata'ita'iga i totonu o vaega fa'apitoa. O lenei faiga e masani ona mana'omia ai le tele o fa'asologa e o'o atu ai i se tulaga maualuga o le sa'o. E masani ona aofia ai le tele o faʻamatalaga e ausia ai le saʻo maualuga, faʻaleleia lelei i luga o galuega eseese ma faʻamaumauga faʻapitoa, faʻamautinoa le faʻatinoga o aʻoaʻoga, ma le saunia o faʻataʻitaʻiga mo le faʻaogaina o faʻamatalaga.

NeMo Framework e maua ai le lagolago mo le aʻoaʻoina ma le faʻavasegaina o faʻataʻitaʻiga ole Speech AI. E aofia ai galuega e pei o le Automatic Speech Recognition (ASR) ma le Text-To-Speech (TTS) synthesis. E ofoina atu le sologa lelei o suiga ile atinaʻe faʻapipiʻi gaosiga faʻatasi ma NVIDIA Riva. Ina ia fesoasoani i tagata atiaʻe ma tagata suʻesuʻe, o le NeMo Framework e aofia ai suʻesuʻega faʻaonaponei aʻo leʻi aʻoaʻoina, meafaigaluega mo le toe gaosia o faʻamatalaga o faʻamatalaga tautala, ma foliga mo suʻesuʻega fefaʻasoaaʻi ma auʻiliʻili o faʻamaumauga o tautalaga. O vaega o le NeMo Framework for Speech AI e faʻapea:

A'oa'oga ma Fa'asinomaga
NeMo Framework o loʻo i ai mea uma e manaʻomia e toleni ma faʻavasegaina faʻataʻitaʻiga tautala (ASR, Fa'avasegaga o tautalaga, Aloaia a le Fofoga Fetalai, Fa'asalalauga Fa'amatalaga, ma TTS) i se faiga e mafai ona toe gaosia.

SOTA Fa'ata'ita'iga na a'oa'oina muamua

• O lo'o tu'uina atu e le NeMo Framework fua fa'aonaponei fa'aonaponei ma nofoaga siaki na mua'i a'oa'oina o nisi ASR ma TTS faʻataʻitaʻiga, faʻapea foʻi ma faʻatonuga i le auala e utaina ai.
• Meafaigaluega Tautala
• NeMo Framework e maua ai se seti o meafaigaluega aoga mo le atinaʻeina o faʻataʻitaʻiga ASR ma TTS, e aofia ai:
  • NeMo Fa'asa'o Fa'asalaina (NFA) mo le fa'atupuina o fa'ailoga, upu-ma vaega-vaega taimiamps o tautalaga i leo e fa'aaoga ai NeMo's CTC-fa'avaeina fa'ata'ita'iga Fa'ailoa Fa'amatalaga Fa'apitoa.
  • Tulaga Fa'amatalaga Fa'amatalaga (SDP), o se meafaigaluega mo le fa'afaigofieina o fa'amaumauga o fa'amatalaga tautala. E fa'atagaina oe e fai ma sui o fa'agaioiga fa'agaioiga o fa'amaumauga i se config file, fa'aiti'itia le code boilerplate ma fa'atagaina le toe gaosia ma le fa'asoa.
  • Su'esu'e Fa'amatalaga Tautala (SDE), o se Dash-fa'avae web talosaga mo su'esu'ega fefa'asoaa'i ma au'ili'ili o fa'amaumauga o fa'amatalaga tautala.
  • Mea faigaluega fau fa'amaumauga lea e maua ai galuega e fa'aoga ai leo uumi files fa'atasi ai ma tusitusiga fa'atatau ma vaevae i ni vaega pupuu e talafeagai mo a'oa'oga fa'ata'ita'iga Fa'ata'ita'iga o le Fa'ailoaina o Tautala Automatic (ASR).
  • Meafaigaluega Faatusatusa mo Fa'ata'ita'iga a le ASR e fa'atusatusa ai valo'aga o fa'ata'ita'iga ASR 'ese'ese i le sa'o o le upu ma le fa'aupuga.
  • Su'esu'e ASR mo le iloiloina o le faatinoga o fa'ata'ita'iga a le ASR ma isi vaega e pei o le Va'aiga Gaoioiga Leo.
  • Meafaigaluega Fa'atonu mo le faaliliuina o tusitusiga mai le pepa tusitusi i le gagana tautala ma le isi itu (eg “31st” vs “thirty first”).
• Auala i le Fa'atulagaina
• O faʻataʻitaʻiga NeMo na aʻoaʻoina pe faʻapitoa e faʻaaoga ai le NeMo Framework e mafai ona faʻaleleia ma faʻapipiʻiina i le NVIDIA Riva. Ua saunia e Riva ni koneteina ma siata Helm ua mamanuina faapitoa e otometi ai laasaga mo le faʻapipiʻiina o le ki.

Isi Punaoa

GitHub Repos

• NeMo: O le fale teu oloa autu mo le NeMo Framework
• NeMo–Taufetuli: O se meafaigaluega e faʻapipiʻi, faʻalauiloa ma pulea au suʻega aʻoaʻoga masini.
• NeMo-Aligner: Meafaigaluega e mafai ona fuaina mo le fa'aogaina lelei o fa'ata'ita'iga
• NeMo-Curator: Fuafuaina o fa'amaumauga a'o le'i fa'agaioiina ma fa'avasega meafaigaluega mo LLM

Mauaina o Fesoasoani
Auai ma le nuu o NeMo, fai fesili, maua se lagolago, pe lipoti mai ni mea sese.

• Talanoaga NeMo
• Mataupu NeMo

Gagana Polokalame ma Fa'ava'a

• Python: Le faʻaoga autu e faʻaaoga ai le NeMo Framework
• Pytorch: NeMo Framework ua fausia i luga ole PyTorch

Laisene

• NeMo Github repo ua laiseneina i lalo ole laisene Apache 2.0
• NeMo Framework ua laiseneina i lalo ole NVIDIA AI PRODUCT AGREEMENT. I le tosoina ma le fa'aogaina o le koneteina, e te talia aiaiga ma aiaiga o lenei laisene.
• O le NeMo Framework container o lo'o iai mea Llama e pulea e le Meta Llama3 Community License Agreement.

vaefaamatalaga
I le taimi nei, NeMo Curator ma NeMo Aligner lagolago mo Multimodal faʻataʻitaʻiga o se galuega o loʻo faʻagasolo ma o le a maua vave.

FAQ

F: E fa'afefea ona ou siaki pe a a'afia la'u faiga ile fa'aletonu?
A: E mafai ona e siaki pe a aʻafia lau polokalama e ala i le faʻamaonia o le faʻaogaina o le NVIDIA NeMo Framework faʻapipiʻiina. Afai o lo'o i lalo ole version 24, e ono a'afia lau faiga.

Q: O ai na lipotia le mataupu saogalemu CVE-2025-23360?
A: O le mataupu tau puipuiga na lipotia mai e Or Peles - JFrog Security. E fa'ailoa e le NVIDIA la latou sao.

F: E fa'afefea ona ou mauaina fa'amatalaga puletini mo le lumana'i?
A: Asiasi i le NVIDIA Product Security itulau e lesitala i faʻamatalaga puletini saogalemu ma nofo logo e uiga i faʻafouga saogalemu o oloa.

Pepa / Punaoa

NVIDIA NeMo Framework [pdf] Taiala mo Tagata Fa'aoga NeMo Framework, NeMo, Framework

Fa'asinomaga

• Tusi Taiala