I-NVIDIA NeMo Framework

Imininingwane

• Igama Lomkhiqizo: I-NVIDIA NeMo Framework
• Izinkundla Ezithintekile: Windows, Linux, macOS
• Izinguqulo Ezithintekile: Zonke izinguqulo zangaphambi kuka-24
• Ukuba sengozini kwezokuvikela: I-CVE-2025-23360
• Isisekelo Sokuhlola Ubungozi: 7.1 (CVSS v3.1)

Imiyalo yokusetshenziswa komkhiqizo

Ukufakwa Kwesibuyekezo Sokuphepha:
Ukuze uvikele isistimu yakho, landela lezi zinyathelo:

1. Landa ukukhishwa kwakamuva okuvela ekhasini lokukhishwa kwe-NeMo-Framework-Launcher ku-GitHub.
2. Iya ku-NVIDIA Product Security ukuze uthole ulwazi olwengeziwe.

Imininingwane Yesibuyekezo Sokuphepha:
Isibuyekezo sokuvikeleka sibhekana nokuba sengozini ku-NVIDIA NeMo Framework okungaholela ekusebenziseni ikhodi kanye nedatha tampukwenza iphutha.

Ukuthuthukiswa Kwesofthiwe:
Uma usebenzisa ukukhululwa kwegatsha kwangaphambilini, kunconywa ukuthi uthuthukele ekukhishweni kwegatsha kwakamuva ukuze kubhekwane nenkinga yezokuphepha.

Kuphelileview

I-NVIDIA NeMo Framework iwuhlaka lwe-AI olwenzekayo noluvela emafini olwakhelwe abacwaningi nabathuthukisi abasebenza kulo. Amamodeli Olimi Amakhulu, Multimodal, kanye Inkulumo AI (isib Ukuqashelwa Kwenkulumo Okuzenzakalelayo futhi Umbhalo-kuya-Enkulumweni). Ivumela abasebenzisi ukuthi bakhe, benze ngokwezifiso, futhi bakhiphe amamodeli amasha e-AI akhiqizayo ngokusebenzisa ikhodi ekhona kanye nezindawo zokuhlola zamamodeli aqeqeshwe kusengaphambili.

Setha Imiyalelo: Faka i-NeMo Framework

Amamodeli Olimi Amakhulu Namamodeli Ahlukahlukene
I-NeMo Framework ihlinzeka ngosekelo olusuka ekupheleni lokuthuthukisa Amamodeli Olimi Olukhulu (LLMs) kanye namamodeli we-Multimodal (MMs). Ihlinzeka ngokuvumelana nezimo ukuze kusetshenziswe endaweni, endaweni yedatha, noma nomhlinzeki wakho wamafu owuncamelayo. Iphinde isekele ukusetshenziswa ku-SLURM noma ezindaweni ezinikwe amandla ze-Kubernetes.

Ukukhethwa Kwedatha
I-NeMo Curator [1] iwumtapo wezincwadi wePython ohlanganisa uchungechunge lwamamojula wokumbiwa kwedatha kanye nokwenza idatha yokwenziwa. Ziyakwazi ukukala futhi zithuthukiselwe ama-GPU, okuwenza alungele ukucutshungulwa idatha yolimi lwemvelo ukuze uqeqeshe noma ucule kahle ama-LLM. Nge-NeMo Curator, ungakwazi ukukhipha umbhalo wekhwalithi ephezulu ngokuningiliziwe okungaphekiwe web imithombo yedatha.

Ukuqeqeshwa nokwenza ngokwezifiso

I-NeMo Framework inikeza amathuluzi okuqeqeshwa okuphumelelayo nokwenza ngokwezifiso Ama-LLM kanye namamodeli we-Multimodal. Kuhlanganisa ukulungiselelwa okuzenzakalelayo kokusetha iqoqo, ukulandwa kwedatha, namapharamitha angamamodeli, angalungiswa ukuze aqeqeshe kumadathasethi amasha namamodeli. Ngokungeziwe ekuqeqeshweni kwangaphambili, i-NeMo isekela kokubili amasu Okucupha Okuhle Okuqondiswayo (i-SFT) kanye ne-Parameter Efficient Fine-Tuning (PEFT) njenge-LoRA, i-Ptuning, nokuningi.

Izinketho ezimbili ziyatholakala ukuze uqalise ukuqeqeshwa ku-NeMo - usebenzisa isixhumi esibonakalayo se-NeMo 2.0 API noma nge-NeMo Run.

• Nge-NeMo Run (Kunconyiwe): I-NeMo Run inikeza isixhumi esibonakalayo ukuze lulaze ukucushwa, ukwenziwa nokuphathwa kokuhlolwa ezindaweni ezihlukahlukene zekhompuyutha. Lokhu kuhlanganisa ukwethula imisebenzi endaweni yakho yokusebenza noma kumaqoqo amakhulu - womabili ama-SLURM anikwe amandla noma i-Kubernetes endaweni yamafu.
  • Ukuqeqeshwa kwangaphambili kanye ne-PEFT Quickstart nge-NeMo Run
• Ukusebenzisa i-NeMo 2.0 API: Le ndlela isebenza kahle ngokusetha okulula okubandakanya amamodeli amancane, noma uma ungathanda ukubhala i-dataloader yakho yangokwezifiso, izihibe zokuqeqesha, noma uguqule izendlalelo zemodeli. Ikunikeza ukuguquguquka okwengeziwe nokulawula ukulungiselelwa, futhi yenza kube lula ukunweba nokwenza ngendlela oyifisayo ukucupha ngokohlelo.
  • I-Training Quickstart nge-NeMo 2.0 API
  • Ifuduka isuka ku-NeMo 1.0 iye ku-NeMo 2.0 API

Ukuqondanisa

• I-NeMo-Aligner [1] iyikhithi yamathuluzi engakala yokuqondisa kahle kwemodeli. Ikhithi yamathuluzi inokusekelwa kwama-algorithms wokuqondanisa amamodeli wesimanje njenge-SteerLM, i-DPO, i-Reinforcement Learning From Human Feedback (RLHF), nokunye okuningi. Lawa ma-algorithms avumela abasebenzisi ukuthi baqondanise amamodeli olimi ukuze aphephe, angabi yingozi, futhi abe wusizo.
• Zonke izindawo zokuhlola ze-NeMo-Aligner ziyahambisana kakhulu ne-NeMo ecosystem, okuvumela ukwenziwa ngokwezifiso okwengeziwe kanye nokusetshenziswa kokuqondisisa.

Ukuhamba komsebenzi kwesinyathelo nesinyathelo kuzo zontathu izigaba ze-RLHF kumodeli encane ye-GPT-2B:

• Ukuqeqeshwa kwe-SFT
• Ukuqeqeshwa kwemodeli yomvuzo
• Ukuqeqeshwa kwe-PPO

Ngaphezu kwalokho, sibonisa ukusekela kwezinye izindlela zokuqondanisa zamanoveli:

• I-DPO: i-algorithm yokuqondanisa engasindi uma iqhathaniswa ne-RLHF enomsebenzi wokulahlekelwa olula.
• Ukuzidlalela Ukushuna Kahle (SPIN)
• I-SteerLM: indlela esuselwe ku-conditioned-SFT, enokuphumayo okuqondisayo.

Bheka amadokhumenti ukuze uthole ulwazi olwengeziwe: Ukuqondanisa Amadokhumenti

Amamodeli we-Multimodal

• I-NeMo Framework inikeza isofthiwe ethuthukisiwe ukuze iqeqeshe futhi ikhiphe amamodeli wesimanjemanje we-multimodal kuzo zonke izigaba ezimbalwa: Amamodeli Olimi Oluningi, Izisekelo Zolimi Lombono, Amamodeli Wombhalo-ube-Isithombe, nangale kwe-2D Generation esebenzisa i-Neural Radiance Fields (NeRF).
• Isigaba ngasinye siklanyelwe ukubhekelela izidingo ezithile kanye nentuthuko emkhakheni, sisebenzisa amamodeli asezingeni eliphezulu ukuphatha uhla olubanzi lwezinhlobo zedatha, okuhlanganisa umbhalo, izithombe, namamodeli e-3D.

Ukusatshalaliswa kanye nencazelo
I-NeMo Framework ihlinzeka ngezindlela ezahlukahlukene zokuchazwa kwe-LLM, ihlinzekela ngezimo ezihlukene zokuthunyelwa kanye nezidingo zokusebenza.

Sebenzisa nge-NVIDIA NIM

• I-NeMo Framework ihlanganisa kalula namathuluzi wokuphakelwa kwemodeli yezinga lebhizinisi nge-NVIDIA NIM. Lokhu kuhlanganiswa kunikwa amandla yi-NVIDIA TensorRT-LLM, iqinisekisa ukuthi inkambiso ithuthukisiwe futhi ingakala.
• Ukuze uthole ulwazi olwengeziwe nge-NIM, vakashela i-NVIDIA webindawo.

Sebenzisa nge-TensorRT-LLM noma i-vLLM

• I-NeMo Framework inikeza imibhalo nama-API ukuze ithekelise amamodeli kumalabhulali amabili athuthukisiwe, i-TensorRT-LLM kanye ne-vLLM, futhi ikhiphe imodeli ethunyelwe nge-NVIDIA Triton Inference Server.
• Ezimweni ezidinga ukusebenza okuthuthukisiwe, amamodeli e-NeMo angasebenzisa i-TensorRT-LLM, umtapo wezincwadi okhethekile wokusheshisa nokuthuthukisa ukuchazwa kwe-LLM kuma-NVIDIA GPU. Le nqubo ibandakanya ukuguqula amamodeli e-NeMo abe ifomethi ehambisana ne-TensorRT-LLM kusetshenziswa imojuli ye-nemo.export.
  • Ukuthunyelwa kwe-LLM Kuphelileview
  • Sebenzisa Amamodeli Olimi Olukhulu lwe-NeMo nge-NIM
  • Sebenzisa Amamodeli Olimi Amakhulu we-NeMo nge-TensorRT-LLM
  • Sebenzisa amamodeli Olimi Amakhulu we-NeMo nge-vLLM

Amamodeli Asekelwe

Amamodeli Olimi Amakhulu

Amamodeli Olimi Amakhulu

Amamodeli Olimi Amakhulu	Ukuqeqeshwa kwangaphambili kanye ne-SFT	I-PEFT	Ukuqondanisa	I-FP8 Training Convergence	I-TRT/TRTLLM	Guqulela Kuye & Ukusuka Ebusweni Obugonileyo	Ukuhlola
Llama3 8B/70B, Llama3.1 405B	Yebo	Yebo	x	Yebo (kuqinisekiswe kancane)	Yebo	Kokubili	Yebo
Ingxubevange 8x7B/8x22B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	Yebo	Kokubili	Yebo
I-Nemotron 3 8B	Yebo	x	x	Yebo (akuqinisekisiwe)	x	Kokubili	Yebo
I-Nemotron 4 340B	Yebo	x	x	Yebo (akuqinisekisiwe)	x	Kokubili	Yebo
I-Baichuan2 7B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	x	Kokubili	Yebo
IngxoxoGLM3 6B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	x	Kokubili	Yebo
I-Gemma 2B/7B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	Yebo	Kokubili	Yebo
I-Gemma2 2B/9B/27B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	x	Kokubili	Yebo
Mamba2 130M/370M/780M/1.3B/2.7B/8B/ Hybrid-8B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	x	x	Yebo
Phi3 mini 4k	x	Yebo	x	Yebo (akuqinisekisiwe)	x	x	x
Qwen2 0.5B/1.5B/7B/72B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	Yebo	Kokubili	Yebo
I-StarCoder 15B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	Yebo	Kokubili	Yebo
I-StarCoder2 3B/7B/15B	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	Yebo	Kokubili	Yebo
I-BERT 110M/340M	Yebo	Yebo	x	Yebo (akuqinisekisiwe)	x	Kokubili	x
T5 220M/3B/11B	Yebo	Yebo	x	x	x	x	x

 

Umbono Amamodeli Olimi

 

Amamodeli Wokushumeka

 

Amamodeli we-World Foundation

Qaphela
I-NeMo futhi isekela ukuqeqeshwa kwangaphambili kukho kokubili ukusabalalisa kanye nezakhiwo ze-autoregressive text2world amamodeli esisekelo.

Inkulumo AI

Ukuthuthukisa amamodeli e-AI yezingxoxo kuyinqubo eyinkimbinkimbi ebandakanya ukuchaza, ukwakha, nokuqeqesha amamodeli ngaphakathi kwezizinda ezithile. Le nqubo ngokuvamile idinga iziphindaphindo ezimbalwa ukuze ifinyelele izinga eliphezulu lokunemba. Kuvame ukubandakanya ukuphindaphinda okuningi ukuze kuzuzwe ukunemba okuphezulu, ukulungisa kahle imisebenzi ehlukahlukene kanye nedatha eqondene nesizinda, ukuqinisekisa ukusebenza kokuqeqeshwa, kanye nokulungiselela amamodeli okuthunyelwa kwezinkomba.

I-NeMo Framework inikeza ukwesekwa kokuqeqeshwa nokwenza ngendlela oyifisayo amamodeli e-Speech AI. Lokhu kufaka phakathi imisebenzi efana ne-Automatic Speech Recognition (ASR) kanye ne-Text-to-Speech (TTS) synthesis. Inikeza uguquko olushelelayo ekusetshenzisweni kokukhiqizwa kwezinga lebhizinisi nge-NVIDIA Riva. Ukusiza abathuthukisi nabacwaningi, i-NeMo Framework ihlanganisa izindawo zokuhlola eziqeqeshwe kusengaphambili, amathuluzi okucubungula idatha yenkulumo ephindaphindekayo, nezici zokuhlola okusebenzisanayo nokuhlaziywa kwamasethi edatha yenkulumo. Izingxenye ze-NeMo Framework for Speech AI zimi kanje:

Ukuqeqeshwa nokwenza ngokwezifiso
I-NeMo Framework iqukethe konke okudingekayo ukuze uqeqeshe futhi wenze ngendlela oyifisayo amamodeli enkulumo (I-ASR, Ukuhlukaniswa Kwenkulumo, Ukuqashelwa Isikhulumi, I-Diarization yesipikha, futhi I-TTS) ngendlela ephindaphindekayo.

Amamodeli Aqeqeshwe Ngaphambi Kwe-SOTA

• I-NeMo Framework inikeza izindlela zokupheka ezisezingeni eliphezulu kanye nezindawo zokuhlola eziqeqeshwe ngaphambilini ezimbalwa I-ASR futhi I-TTS amamodeli, kanye neziyalezo zendlela yokulayisha.
• Amathuluzi Okukhuluma
• I-NeMo Framework inikeza isethi yamathuluzi awusizo ekuthuthukiseni amamodeli e-ASR kanye ne-TTS, okuhlanganisa:
  • I-NeMo Forced Aligner (NFA) ukuze kukhiqizwe ithokheni-, igama- kanye nesikhathi seleveli yengxenyeampyenkulumo emsindweni kusetshenziswa amamodeli e-NeMo's CTC-based Automatic Speech Recognition.
  • I-Speech Data Processor (SDP), ikhithi yamathuluzi yokwenza lula ukucubungula idatha yenkulumo. Ikuvumela ukuthi umelele imisebenzi yokucubungula idatha ku-config file, ukunciphisa ikhodi ye-boilerplate kanye nokuvumela ukuphindaphinda nokwabelana.
  • I-Speech Data Explorer (SDE), i-Dash-based web isicelo sokuhlola okusebenzisanayo nokuhlaziywa kwamasethi edatha yenkulumo.
  • Ithuluzi lokudala isethi yedatha okunikeza ukusebenza ukuvumelanisa umsindo omude files nemibhalo ebhaliwe ehambisanayo futhi ihlukaniseke ibe yizingcezu ezimfushane ezifanele ukuqeqeshwa kwemodeli ye-Automatic Speech Recognition (ASR).
  • Ithuluzi lokuqhathanisa ukuze Amamodeli we-ASR aqhathanise ukuqagela kwamamodeli ahlukene e-ASR ekunembeni kwamagama nezinga lokuphimisa.
  • Umhloli we-ASR ukuze kuhlolwe ukusebenza kwamamodeli we-ASR nezinye izici ezifana nokutholwa komsebenzi wezwi.
  • Ithuluzi Lokwenza Umbhalo Ojwayelekile ukuguqula umbhalo usuke endleleni ebhaliwe uyiswe okhulumayo futhi okuphambene nalokho (isb. “yama-31” vs “amashumi amathathu nanye”).
• Indlela eya Ekusetshenzisweni
• Amamodeli e-NeMo aqeqeshiwe noma enziwe ngendlela oyifisayo kusetshenziswa i-NeMo Framework angathuthukiswa futhi asetshenziswe ne-NVIDIA Riva. I-Riva ihlinzeka ngeziqukathi namashadi e-Helm aklanyelwe ngokuqondile ukwenza izinyathelo zokuthunyelwa kwenkinobho yokuphusha.

Ezinye Izinsiza

I-GitHub Repos

• I-NeMo: Inqolobane eyinhloko ye-NeMo Framework
• I-NeMo-Gijima: Ithuluzi lokumisa, ukuqalisa nokuphatha ukuhlolwa komshini wakho wokufunda.
• I-NeMo-Aligner: Ikhithi yamathuluzi ekalayo ukuze imodeli iqondaniswe kahle
• I-NeMo-Curator: Ukucutshungulwa kwangaphambili kwedatha kanye nekhithi yamathuluzi ehlunga ama-LLM

Ukuthola Usizo
Xhumana nomphakathi we-NeMo, buza imibuzo, thola ukwesekwa, noma ubike iziphazamisi.

• Izingxoxo ze-NeMo
• Izinkinga ze-NeMo

Izilimi Zokuhlela kanye Nezinhlaka

• I-Python: Isixhumi esibonakalayo esiyinhloko sokusebenzisa i-NeMo Framework
• I-Pytorch: I-NeMo Framework yakhelwe phezu kwe-PyTorch

Amalayisense

• I-NeMo Github repo ilayisensi ngaphansi kwelayisense ye-Apache 2.0
• I-NeMo Framework inikezwe ilayisense ngaphansi kwe-NVIDIA AI PRODUCT AGREEMENT. Ngokudonsa nokusebenzisa isiqukathi, uyayamukela imigomo nemibandela yale layisensi.
• Isiqukathi se-NeMo Framework siqukethe izinto ze-Llama ezibuswa Isivumelwano Selayisensi Yomphakathi ye-Meta Llama3.

Imibhalo yaphansi
Njengamanje, ukwesekwa kwe-NeMo Curator kanye ne-NeMo Aligner yamamodeli we-Multimodal kuwumsebenzi oqhubekayo futhi uzotholakala maduze.

FAQ

Q: Ngingahlola kanjani ukuthi isistimu yami ithintwa ukuba sengozini?
IMP: Ungahlola ukuthi isistimu yakho iyathinteka yini ngokuqinisekisa inguqulo ye-NVIDIA NeMo Framework efakiwe. Uma ingaphansi kwenguqulo 24, isistimu yakho ingaba sengozini.

Q: Ubani obike udaba lwezokuphepha CVE-2025-23360?
IMP: Indaba yezokuphepha ibikwe ngabakwa-Or Peles – JFrog Security. I-NVIDIA iyawazisa umnikelo wabo.

Q: Ngingazithola kanjani izaziso zezindaba zokuphepha zesikhathi esizayo?
A: Vakashela ikhasi le-NVIDIA Product Security ukuze ubhalisele izaziso zezindaba zokuphepha futhi uhlale unolwazi mayelana nezibuyekezo zokuphepha komkhiqizo.

Amadokhumenti / Izinsiza

I-NVIDIA NeMo Framework [pdf] Umhlahlandlela Womsebenzisi I-NeMo Framework, i-NeMo, i-Framework

Izithenjwa

• Imaniwali yosebenzisayo