NVIDIA NeMo Framework
Zvinotsanangurwa
- Product Name: NVIDIA NeMo Framework
- Mapuratifomu Akanganiswa: Windows, Linux, macOS
- Affected Versions: Shanduro dzese dzisati dzasvika 24
- Chengetedzo Dambudziko: CVE-2025-23360
- Risk Assessment Base Score: 7.1 (CVSS v3.1)
Mirayiridzo Yekushandiswa Kwechigadzirwa
Security Update Installation:
Kuti udzivirire system yako, tevera matanho aya:
- Dhawunirodha yazvino kuburitswa kubva kuNeMo-Framework-Launcher Releases peji paGitHub.
- Enda kuNVIDIA Chigadzirwa Chekuchengetedza kuti uwane rumwe ruzivo.
Security Update Details:
Iyo yekuchengetedza yekuvandudza inogadzirisa kusagadzikana muNVIDIA NeMo Framework iyo inogona kutungamira mukuita kodhi uye data t.ampering.
Software Ndiwedzere:
Kana uri kushandisa kuburitswa kwebazi kwekare, zvinokurudzirwa kukwidziridza kune yazvino kuburitswa kwebazi kugadzirisa dambudziko rekuchengetedza.
Overview
NVIDIA NeMo Framework ndeye scalable uye gore-yekuzvarwa generative AI chimiro chakavakirwa vaongorori nevagadziri vanoshanda pa Mienzaniso Yakakura MutauroMultimodal, uye Kutaura AI (semuenzaniso Otomatiki Kuzivikanwa Kwekutaura uye Chinyorwa-ku-Kutaura) Inogonesa vashandisi kugadzira, kugadzirisa, uye kuendesa mhando nyowani dzeAI nekusimudzira kodhi iripo uye pre-yakadzidziswa modhi yekutarisa.
Setup Instructions: Isa NeMo Framework
NeMo Framework inopa magumo-kumagumo rutsigiro rwekugadzira Makuru Mutauro Models (LLMs) uye Multimodal Models (MMs). Inopa kuchinjika kushandiswa pa-nzvimbo, mune data-center, kana neyako yaunofarira gore mupi. Inotsigirawo kuurayiwa paSLURM kana Kubernetes inogonesa nharaunda.
Data Curation
NeMo Curator [1] iraibhurari yePython inosanganisira suite yemamodule ekuchera data uye kugadzira data kugadzira. Iwo ane scalable uye akagadziridzwa maGPU, achiaita akanakira kudzora dhata remutauro wechisikigo kudzidzisa kana kunyatso-tuna maLLM. NeNeMo Curator, unogona kunyatso tora mavara emhando yepamusoro kubva kune yakasvibira web data sources.
Kudzidzira uye Customization
NeMo Framework inopa maturusi ekudzidzisa kwakanaka uye kugadzirisa LLMs uye Multimodal modhi. Inosanganisira zvigadziriso zvekugadzirisa komputa cluster setup, kudhawunirodha data, uye modhi hyperparameters, inogona kugadziridzwa kudzidzisa pane nyowani dataset uye modhi. Pamusoro pekutanga kudzidziswa, NeMo inotsigira ese akatariswa Yakanatswa-Tuning (SFT) uye Parameter Inoshanda Yakanaka-Tuning (PEFT) maitiro seLoRA, Ptuning, nezvimwe.
Sarudzo mbiri dziripo kuti utange kudzidziswa muNeMo - uchishandisa NeMo 2.0 API interface kana neNeMo Run.
- NeNeMo Run (Inokurudzirwa): NeMo Run inopa chinongedzo chekugadzirisa zvigadziriso, kuita uye manejimendi ezviedzo munzvimbo dzakasiyana siyana dzekombuta. Izvi zvinosanganisira kutangisa mabasa panzvimbo yako yekushandira munharaunda kana pamasumbu makuru - ese SLURM akagoneswa kana Kubernetes munzvimbo yegore.
- Pre-kudzidziswa & PEFT Quickstart neNeMo Run
- Kushandisa iyo NeMo 2.0 API: Iyi nzira inoshanda zvakanaka neyakapfava setup inosanganisira madiki mamodheru, kana iwe uchifarira kunyora yako wega tsika dataloader, kudzidzira zvishwe, kana kushandura modhi layer. Iyo inokupa iwe kuchinjika uye kutonga pamusoro pezvigadziriso, uye inoita kuti zvive nyore kuwedzera uye kugadzirisa zvigadziriso zvine hurongwa.
-
Training Quickstart neNeMo 2.0 API
-
Kutama kubva kuNeMo 1.0 kuenda kuNeMo 2.0 API
-
Alignment
- NeMo-Aligner [1] ndeye scalable toolkit yekumisikidza modhi yakanaka. The toolkit ine tsigiro yemamiriro-of-the-art model alignment algorithms akadai seSteerLM, DPO, Reinforcement Learning kubva kuHuman Feedback (RLHF), nezvimwe zvakawanda. Aya maalgorithms anogonesa vashandisi kuwiriranisa mamodheru emitauro kuti ive yakachengeteka, isingakuvadze, uye inobatsira.
- Ese maNeMo-Aligner ekutarisa anoyambuka-anoenderana neNeMo ecosystem, achibvumira kuenderera mberi nekugadzirisa uye inference deployment.
Nhanho-ne-nhanho mafambiro ezvikamu zvese zvitatu zveRLHF pane diki GPT-2B modhi:
- SFT kudzidziswa
- Mubayiro muenzaniso kudzidziswa
- PPO kudzidziswa
Pamusoro pezvo, isu tinoratidza tsigiro yedzimwe nzira dzakasiyana dzekurongedza:
- DPO: algorithm yakareruka yakaenzaniswa neRLHF ine basa rakareruka rekurasikirwa.
- Kuzvitamba Kunyatsogadzirisa (SPIN)
- SteerLM: hunyanzvi hwakavakirwa pane conditioned-SFT, ine steerable inobuda.
Tarisa magwaro kuti uwane rumwe ruzivo: Alignment Documentation
Multimodal Models
- NeMo Framework inopa yakagadziridzwa software kudzidzisa uye kuendesa mamiriro-e-the-art multimodal modhi muzvikamu zvakati wandei: Multimodal Mutauro Models, Vision-Mutauro Nheyo, Zvinyorwa-ku-Image modhi, uye kupfuura 2D Generation uchishandisa Neural Radiance Fields (NeRF).
- Chikamu chega chega chakagadzirirwa kuenderana nezvinodiwa uye kufambira mberi mumunda, kusimudzira mhando dzekucheka-kumucheto kubata dzakasiyana siyana dzemhando dzedata, kusanganisira zvinyorwa, mifananidzo, uye 3D modhi.
Cherechedza
Isu tiri kutama rutsigiro rwemhando dzakawanda kubva kuNeMo 1.0 kuenda kuNeMo 2.0. Kana iwe uchida kuongorora iyi domain panguva ino, ndapota tarisa kune zvinyorwa zveNeMo 24.07 (yapfuura) kuburitswa.
Deployment uye Inference
NeMo Framework inopa nzira dzakasiyana dzeLLM inference, inobata kune akasiyana ekuisa mamiriro uye kuita zvinodiwa.
Shandisa neNVIDIA NIM
- NeMo Framework inobatanidza zvisina mutsetse nemabhizimusi-level modhi yekuisa maturusi kuburikidza neNVIDIA NIM. Kubatanidzwa uku kunopihwa simba neNVIDIA TensorRT-LLM, kuve nechokwadi chakagadziridzwa uye chinotyisa inference.
- Kuti uwane rumwe ruzivo nezveNIM, shanyira iyo NVIDIA website.
Shandisa neTensorRT-LLM kana vLLM
- NeMo Framework inopa zvinyorwa uye APIs kutumira modhi kune maviri inference optimized raibhurari, TensorRT-LLM uye vLLM, uye kuendesa iyo inotengeswa modhi neNVIDIA Triton Inference Server.
- Kune mamiriro ezvinhu anoda optimized performance, NeMo modhi inogona kukwirisa TensorRT-LLM, raibhurari yakasarudzika yekumhanyisa uye nekugadzirisa LLM inference paNVIDIA GPUs. Kuita uku kunosanganisira kushandura mhando dzeNeMo kuita fomati inoenderana neTensorRT-LLM uchishandisa nemo.export module.
- LLM Deployment Overview
- Shandisa NeMo Makuru Mutauro Models neNIM
- Shandisa NeMo Makuru Mutauro Models neTensorRT-LLM
- Shandisa NeMo Makuru Mutauro Models ane vLLM
Inotsigirwa Mienzaniso
Mienzaniso Yakakura Mutauro
| Mienzaniso Yakakura Mutauro | Pretraining & SFT | PEFT | Alignment | FP8 Kudzidzisa Convergence | TRT/TRTLLM | Shandura Ku & Kubva Kumbundira Chiso | Evaluation |
|---|---|---|---|---|---|---|---|
| Llama3 8B/70B, Llama3.1 405B | Ehe | Ehe | x | Hongu (zvakasimbiswa zvishoma) | Ehe | Zvose | Ehe |
| Mixtral 8x7B/8x22B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | Ehe | Zvose | Ehe |
| Nemotron 3 8B | Ehe | x | x | Hongu (hazvina kusimbiswa) | x | Zvose | Ehe |
| Nemotron 4 340B | Ehe | x | x | Hongu (hazvina kusimbiswa) | x | Zvose | Ehe |
| Baichuan2 7B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Zvose | Ehe |
| ChatGLM3 6B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Zvose | Ehe |
| Gemma 2B/7B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | Ehe | Zvose | Ehe |
| Gemma2 2B/9B/27B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Zvose | Ehe |
| Mamba2 130M/370M/780M/1.3B/2.7B/8B/ Hybrid-8B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | x | Ehe |
| Phi3 mini 4k | x | Ehe | x | Hongu (hazvina kusimbiswa) | x | x | x |
| Qwen2 0.5B/1.5B/7B/72B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | Ehe | Zvose | Ehe |
| StarCoder 15B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | Ehe | Zvose | Ehe |
| StarCoder2 3B/7B/15B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | Ehe | Zvose | Ehe |
| BERT 110M/340M | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Zvose | x |
| T5 220M/3B/11B | Ehe | Ehe | x | x | x | x | x |
Vision Language Models
| Vision Language Models | Pretraining & SFT | PEFT | Alignment | FP8 Kudzidzisa Convergence | TRT/TRTLLM | Shandura Ku & Kubva Kumbundira Chiso | Evaluation |
|---|---|---|---|---|---|---|---|
| NeVA (LLaVA 1.5) | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Kubva | x |
| Llama 3.2 Vision 11B/90B | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Kubva | x |
| LLaVA Inotevera (LLaVA 1.6) | Ehe | Ehe | x | Hongu (hazvina kusimbiswa) | x | Kubva | x |
Embedding Models
| Kupinza Mienzaniso Yemitauro | Pretraining & SFT | PEFT | Alignment | FP8 Kudzidzisa Convergence | TRT/TRTLLM | Shandura Ku & Kubva Kumbundira Chiso | Evaluation |
|---|---|---|---|---|---|---|---|
| SBERT 340M | Ehe | x | x | Hongu (hazvina kusimbiswa) | x | Zvose | x |
| Llama 3.2 Kupinza 1B | Ehe | x | x | Hongu (hazvina kusimbiswa) | x | Zvose | x |
World Foundation Models
| World Foundation Models | Post-Kudzidziswa | Accelerated Inference |
|---|---|---|
| Cosmos-1.0-Diffusion-Text2World-7B | Ehe | Ehe |
| Cosmos-1.0-Diffusion-Text2World-14B | Ehe | Ehe |
| Cosmos-1.0-Diffusion-Video2World-7B | Kuuya manje | Kuuya manje |
| Cosmos-1.0-Diffusion-Video2World-14B | Kuuya manje | Kuuya manje |
| Cosmos-1.0-Autoregressive-4B | Ehe | Ehe |
| Cosmos-1.0-Autoregressive-Video2World-5B | Kuuya manje | Kuuya manje |
| Cosmos-1.0-Autoregressive-12B | Ehe | Ehe |
| Cosmos-1.0-Autoregressive-Video2World-13B | Kuuya manje | Kuuya manje |
Cherechedza
NeMo inotsigirawo pretraining kune ese ari maviri diffusion uye autoregressive architecture text2world nheyo mienzaniso.
Kutaura AI
Kugadzira yekukurukurirana AI modhi inzira yakaoma inosanganisira kutsanangura, kuvaka, uye kudzidzisa modhi mukati memamwe madomasi. Iyi nzira inowanzoda kudzokororwa kwakati kuti isvike padanho repamusoro rekurongeka. Inowanzo zvinosanganisira kudzokororwa kwakawanda kuti uwane kurongeka kwepamusoro, kunyatsogadzirisa pamabasa akasiyana-siyana uye dhata-chaiyo dhata, kuve nechokwadi chekudzidzira kuita, uye kugadzirira modhi dzekutumira.
NeMo Framework inopa rutsigiro rwekudzidziswa uye kugadzirisa kweKutaura AI modhi. Izvi zvinosanganisira mabasa akaita seAutomatic Speech Recognition (ASR) uye Text-to-Speech (TTS) synthesis. Inopa shanduko yakatsetseka kune bhizinesi-chikamu chekugadzira dhizaini neNVIDIA Riva. Kubatsira vanogadzira uye vanotsvaga, NeMo Framework inosanganisira mamiriro-e-the-art pre-yakadzidziswa yekutarisa, maturusi ekudhirowazve ekutaura dhata kugadzirisa, uye maficha ekuongorora anopindirana uye kuongororwa kwemaseti ekutaura. Izvo zvikamu zveNeMo Framework yeKutaura AI ndizvo zvinotevera:
Kudzidzira uye Customization
NeMo Framework ine zvese zvinodiwa kudzidzisa uye kugadzirisa matauriro emhando (ASR, Kuronga Kwekutaura, Kuzivikanwa kweMutauriri, Mutauri Diarization,uye TTS) nenzira inogoneka.
SOTA Pre-kudzidziswa Models
- NeMo Framework inopa mamiriro-e-the-art mabikirwo uye pre-yakadzidziswa yekutarisa yeanoverengeka ASR uye TTS mamodheru, pamwe nemirayiridzo yekuti ungaaisa sei.
- Zvishandiso Pakutaura
- NeMo Framework inopa seti yezvishandiso zvinobatsira kugadzira ASR neTTS modhi, kusanganisira:
- NeMo Forced Aligner (NFA) yekugadzira chiratidzo-, izwi- uye chikamu-chikamu nguvaamps yekutaura muodhiyo uchishandisa NeMo's CTC-yakavakirwa Otomatiki Kutaura Kuzivikanwa modhi.
- Speech Data Processor (SDP), chishandiso chekurerutsa magadzirirwo edata rekutaura. Iyo inokutendera kuti umiririre data data mashandiro mune config file, kuderedza boilerplate kodhi uye kubvumira kuberekana uye kugovana.
- Speech Data Explorer (SDE), a Dash-based web application yekuongorora inopindirana uye ongororo yemaseti ekutaura.
- Dataset kugadzira chishandiso iyo inopa mashandiro ekubatanidza kureba odhiyo files ine zvinyorwa zvinoenderana uye wozvipatsanura kuita zvidimbu zvipfupi zvinokodzera Automatic Speech Recognition (ASR) modhi yekudzidziswa.
- Kuenzanisa Tool yeASR Models kuenzanisa kufanotaura kwemhando dzakasiyana dze ASR pakutaura kweshoko uye padanho rekutaura.
- ASR Muongorori yekuongorora mashandiro eASR modhi uye zvimwe zvakaita seVoice Activity Detection.
- Chinyorwa Normalization Tool yekushandura zvinyorwa kubva pakunyora kuenda kuzita rekutaura zvichipesana (eg "yemakumi matatu nerimwe" vs "makumi matatu nekutanga").
- Nzira Yekuendeswa
- NeMo modhi dzakadzidziswa kana kugadziridzwa uchishandisa NeMo Framework inogona kuvandudzwa uye kuiswa pamwe neNVIDIA Riva. Riva inopa midziyo uye Helm machati akagadzirirwa zvakananga kuti aite otomatiki matanho ekusundidzira-bhatani kuendesa.
Zvimwe Zvishandiso
- NeMo: Iyo huru repository yeNeMo Framework
- NeMo-Mhanyai: Chishandiso chekugadzirisa, kuvhura uye kugadzirisa zviedzo zvako zvekudzidza muchina.
- NeMo-Aligner: Scalable toolkit yekuenzanisa muenzaniso wakanaka
- NeMo-Curator: Scalable data pre-processing uye curation toolkit yeLLMs
Bata nenharaunda yeNeMo, bvunza mibvunzo, wana rutsigiro, kana shuma tsikidzi.
- NeMo Hurukuro
- NeMo Issues
Programming Mitauro uye Zvirongwa
- Python: Iyo huru interface yekushandisa NeMo Framework
- Pytorch: NeMo Framework yakavakwa pamusoro pePyTorch
Marezenisi
- NeMo Github repo ine rezinesi pasi peApache 2.0 rezinesi
- NeMo Framework ine rezinesi pasi peNVIDIA AI PRODUCT AGREEMENT. Nekudhonza uye kushandisa mudziyo, unobvuma zviga nemamiriro erezinesi iri.
- Iyo NeMo Framework mudziyo ine Llama zvinhu zvinotongwa neMeta Llama3 Community License Agreement.
Mashoko omuzasi
Parizvino, NeMo Curator uye NeMo Aligner rutsigiro rweMultimodal modhi ibasa ririkuitika uye richavepo munguva pfupi iri kutevera.
FAQ
Q: Ndingatarise sei kana system yangu yakakanganiswa nekusagadzikana?
A: Unogona kutarisa kana sisitimu yako yakanganiswa nekuona iyo vhezheni yeNVIDIA NeMo Framework yakaiswa. Kana iri pazasi vhezheni 24, system yako inogona kuve panjodzi.
Mubvunzo: Ndiani akashuma nyaya yekuchengetedza CVE-2025-23360?
A: Nyaya yekuchengetedza yakataurwa naOr Peles - JFrog Security. NVIDIA inobvuma kubatsira kwavo.
Mubvunzo: Ndingagashira sei zviziviso zvekuchengetedzwa kweramangwana?
A: Shanyira iyo NVIDIA Chigadzirwa Chekuchengetedza peji kuti unyore kune chengetedzo bulletin zviziviso uye ugare uine ruzivo nezve zvigadzirwa zvekuchengetedza zvigadzirwa.
Zvinyorwa / Zvishandiso
 |
NVIDIA NeMo Framework [pdf] Bhuku reMushandisi NeMo Framework, NeMo, Framework |