Optimiser bi dafay wàcci ci CPU ak NVMe
Ab pexe buy sakkanal mémoire biy gaare ay téere yu diis ci tàggat (statu optimiser, gradient, yenn saa yi diisaay) ci RAM CPU wala ci NVMe SSDs ci barabu mémoire GPU bu néew.
Résumé
It lets people train far larger models than their GPU's memory would otherwise allow.
Plongeur bu xóot
Soo tàggatee benn reso neuronal ak optimisatër bu melni Adam, parametre bu nekk dafay yóbbaale bagaas bu gëna bari: ñaari lim yuy daw (momentum ak variance), boole ci benn kopi bu leer ci diisaay bi, boole ci gradient bi. Ci tàggat yaram bu jaxaso, loolu mën na nekk lu tollu ci 16 byte ci parametre bu nekk, di wàññi 2 byte yi ci diisaay bi ci boppam. Offloading dafay dindi bagaas bi ci GPU bi. Decharge CPU dafay génne stade optimiser yi ci sistem RAM bu gëna neex ci kaw bus PCIe, fekk NVMe offload daf leen di puus ba ci disk solid-state yu gaaw yi. ZeRO-Infinity ak ZeRO-Offload yu DeepSpeed ñoo ko siiwal, pexem dafay wecci gaawaay bu ñor ak kàttan, may benn GPU wala cluster bu ndaw ñu defar ay model yu am ay miliyaar ciy parametre.
Gis-gis xarala
Caabi gi mooy boole done yi ak xayma yi. Optimisatër bi dafa toog ci CPU/NVMe; ci jamonoy dellu ginaaw, dañuy njëkka jël xaaj yi ci kaw PCIe balaa ñu leen di soxla, te jéego optimiser bi ci boppam dafay faral di dox ci CPU bi. ZeRO-Offload dafay tëye poids master float32 ak Adam moments ci CPU, kon math ci kanam ak ci ginaaw kese lañuy des ci GPU bi. NVMe dafay yokk cache bu bari suko defee etaa yu terabyte yi di tuuru ci disk bi fekk xaaj yu tàng yi des ci RAM bi.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu Optimizer biy wàcci ci CPU ak NVMe
Ginaaw model yi dañuy wéy di gëna màgg ci mémoire GPU, dechargement bu bari dafay nekk standard te baña nekk lu wuute. Xaarandil lëkkaloo bu gëna dëgër ak lëkkaloo yu gëna gaaw yu melni NVLink-C2C ak CXL pool memory yuy nëbb CPU-GPU, boole ci ay scheduler yu gëna am xel yuy wax ban etaa lañu wara prefetch. Architecture yu am mémoire buñ boole mu melni Grace Hopper dañuy wàññi penalti PCIe, te kaadar yi dañuy puus ngir def offload bu bari daanaka leer nàññ suko defee ñi bëgg bëgg mën nañu defar model yu mag ci hardware bu woyof.
Doxal ci àdduna dëgg
Defar ab LLM bu am 13 milyaar ci benn GPU bu 24 GB di jëfandikoo DeepSpeed ZeRO-Offload ngir push Adam ci RAM CPU bi.
Laboratoire gëstu bu ndaw buy tàggat xeetu parametre yu bari ci yenn GPUs ci di tuuru etaa optimisatër yi ci NVMe drive ak ZeRO-Infinity.
Hugging Face Gaawaay configs yiy may CPU offload suko defee jëfandikukat yi mëna def liggéey yu mat sëkk yu mën sànni njuumti yi ci memory bi.
New-conscious startups yi luwaase GPUs cloud yu gëna xéewale, memory bu néew, ak yobbu ko ci NVMe biñ takk ci barabu fay kàrt 80 GB yu gëna kawe.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Optimizer State Offloading to CPU and NVMe quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Adam ak Optimisateurs Adaptifs
Laaj yi ñuy faral di laaj
What is Optimizer State Offloading to CPU and NVMe?
Ab pexe buy sakkanal mémoire biy gaare ay téere yu diis ci tàggat (statu optimiser, gradient, yenn saa yi diisaay) ci RAM CPU wala ci NVMe SSDs ci barabu mémoire GPU bu néew. Dafay may nit ñi ñu tàggat model yu gëna mag te seen mémoire GPU mënu ko woon may.
Lan mooy mébet bu njëkk bi ñuy dindi etaa optimiser yi ci CPU wala NVMe?
Offloading dafay toxal xeetu optimizer yu diis yi ci GPU bi ci CPU RAM wala NVMe, mooy dindi mémoire GPU bi suko defee ñu mëna tàggat model yu gëna mag ci benn hardware bi.
Ngir optimisatëru Adam ci njubte bu jaxaso, ban done mooy jëfandikoo memory bi gëna bari ci parametre bu nekk?
Adam dafay denc momentum, variance, ak poid master bu mat sëkk, lu tollu ci 16 byte ci parametre bu nekk, lu ëpp bu baax poid bu 2 byte ci xaaju presion.
Ban kaadar moo siiwal dechargement optimiser bu mag?
ZeRO-Offload ak ZeRO-Infinity yu DeepSpeed dugal nañu ba noppi siiwal réew yu baax yi ci CPU ak NVMe ci eskaal bi.
Lan mooy njëgu performance bi gëna mag ci wàcce ci CPU wala NVMe?
Toxal etaa yi ci biti GPU mooy toxal done ci PCIe wala ci NVMe, lu gëna néew memory ci GPU, kon produit bi dafay wàññeeku fileek jaxasoowul ak ordinatër.
naka la sistemu decharge di nëbbe lu bari ci latency transfert bi?
Programer yi dañuy njëkka jël partition yiñ soxla ci PCIe bi GPU bi di ordinatër, di jaxasoo jokkoo ak ordinatër ngir maske latency.