I-OpenXLA, iprojekthi yomthombo ovulekileyo wokukhawulezisa kunye nokwenza lula ukufundwa koomatshini

Mva nje ezona nkampani zinkulu zibandakanyeka kuphuhliso kwinkalo yokufunda koomatshini ebonisiweyo iprojekthi OpenXLA, ejoliswe kuphuhliso oludibeneyo lwezixhobo ukuqokelela kunye nokwenza ngcono iimodeli zeenkqubo zokufunda ngoomatshini.

Iprojekthi ithathe uxanduva lophuhliso lwezixhobo ezivumela ukudibanisa ukuqulunqwa kweemodeli ezilungiselelwe kwi-TensorFlow, i-PyTorch kunye nesikhokelo se-JAX soqeqesho olusebenzayo kunye nokuphunyezwa kwii-GPU ezahlukeneyo kunye nee-accelerator ezikhethekileyo. Iinkampani ezifana neGoogle, NVIDIA, AMD, Intel, Meta, Apple, Arm, Alibaba kunye ne-Amazon bajoyine umsebenzi odibeneyo weprojekthi.

IProjekthi ye-OpenXLA inikezela nge-ML compiler ye-ML enokulinganisa phakathi kobunzima beziseko ze-ML. Iintsika zayo ezisisiseko kukusebenza, ukukala, ukuphatheka, ukuguquguquka kunye nokwandiswa kwabasebenzisi. Nge-OpenXLA, sinqwenela ukuvula amandla okwenyani e-AI ngokukhawulezisa uphuhliso kunye nokuhanjiswa kwayo.

I-OpenXLA yenza abaphuhlisi bakwazi ukuqulunqa kunye nokwandisa iimodeli kuzo zonke izikhokelo zeML ezikhokelayo zoqeqesho olusebenzayo kunye nenkonzo kwiintlobo ngeentlobo zehardware. Abaphuhlisi abasebenzisa i-OpenXLA baya kubona ukuphucuka okubalulekileyo kwixesha loqeqesho, ukusebenza, ukubambezeleka kwenkonzo, kwaye ekugqibeleni ixesha lokuthengisa kunye nokubala iindleko.

Kuyathenjwa ukuba ngokujoyina iinzame yamaqela aphambili ophando kunye nabameli boluntu, kuya kwenzeka ukukhuthaza uphuhliso lweenkqubo zokufunda koomatshini kunye nokusombulula ingxaki yokwahlulwa kweziseko zophuhliso kwiinkqubo ezahlukeneyo namaqela.

I-OpenXLA ivumela ukuphumeza inkxaso esebenzayo kwii-hardware ezahlukeneyo, kungakhathaliseki ukuba yeyiphi inkqubo esekelwe kuyo imodeli yokufunda umatshini. I-OpenXLA kulindeleke ukuba inciphise ixesha loqeqesho lwemodeli, iphucule ukusebenza, inciphise i-latency, inciphise i-computing overhead, kwaye inciphise ixesha lokuthengisa.

OpenXLA iqulathe amacandelo amathathu aphambili, ikhowudi leyo isasazwe phantsi kwelayisensi ye-Apache 2.0:

1. I-XLA (i-algebra ekhawulezileyo yomgca) ngumqokeleli okuvumela ukuba ukhulise iimodeli zokufunda zoomatshini ukuze usebenze ngokusebenza okuphezulu kumaqonga ohlukeneyo ehardware, kuquka ii-GPU, ii-CPU, kunye nee-accelerator ezikhethekileyo ezivela kubavelisi abohlukeneyo.
2. I-StableHLO luphawu olusisiseko kunye nokuphunyezwa kweseti yeMisebenzi yeNqanaba eliPhezulu (HLOs) ukuze isetyenziswe kwiimodeli zenkqubo yokufunda koomatshini. Isebenza njengomaleko phakathi kwezicwangciso zokufunda zoomatshini kunye nabaqulunqi abaguqula imodeli ukuba isebenze kwi-hardware ethile. Iileya zilungiselelwe ukuvelisa imifuziselo kwifomathi yeStableHLO yePyTorch, TensorFlow kunye nesakhelo seJAX. I-MHLO suite isetyenziswa njengesiseko se-StableHLO, eyandisiweyo ngenkxaso yolandelelwano kunye nolawulo loguqulelo.
3. I-IREE (Imeko ePhakathi yokuMelelwa kokuSebenza) ngumqambi kunye nexesha lokuqhuba eliguqula iimodeli zokufunda zoomatshini zibe ngumelo ophakathi jikelele olusekelwe kwi-MLIR (i-Intermediate Multi-Level Representation) yeprojekthi ye-LLVM. Kwiimpawu, ukubakho kokuhlanganiswa kwangaphambili (phambi kwexesha), inkxaso yolawulo lokuhamba, ukukwazi ukusebenzisa izinto eziguquguqukayo kwiimodeli, ukulungelelaniswa kwee-CPU ezahlukeneyo kunye nee-GPU, kunye nentloko ephantsi igxininisiwe.

Ngokumalunga nezibonelelo eziphambili ze-OpenXLA, kukhankanyiwe ukuba ukusebenza ngokugqibeleleyo kuphunyeziwe ngaphandle kokungena kwikhowudi yokubhala isixhobo esithile, ukongeza kwi bonelela ngaphandle kwebhokisi ulungiselelo, kubandakanywa ukucaciswa kwe-algebraic expressions, ukwabiwa kwememori ngokufanelekileyo, ukucwangciswa kokwenziwa, kuthathelwa ingqalelo ukunciphisa ukusetyenziswa kwememori ephezulu kunye nokugqithisa.

Enye inzuzo yi ukwenziwa lula kokulinganisa kunye nokulungelelanisa izibalo. Kwanele ukuba umphuhlisi wongeze ama-annotations kwi-subset ye-tensors ebalulekileyo, ngesiseko apho umqambi angenza ngokuzenzekelayo ikhowudi ye-computing ehambelanayo.

Kuyacaciswa ukuba ukuphatheka kubonelelwa ngenkxaso yamaqonga ehardware amaninzi, njenge-AMD kunye ne-NVIDIA GPUs, i-x86 kunye nee-ARM CPU, i-Google TPU ML Accelerators, i-AWS Trainium Inferentia IPUs, iGraphcore, kunye ne-Wafer-Scale Engine Cerebras.

Inkxaso yokudibanisa izandiso kunye nokuphunyezwa kwemisebenzi eyongezelelweyo, njengenkxaso yokubhala iiprimitives zokufunda ezinzulu zoomatshini usebenzisa iCUDA, HIP, SYCL, Triton kunye nezinye iilwimi kwikhompyuter efanayo, kunye ukubakho kohlengahlengiso lwezandla lweebhotile kwiimodeli.

Okokugqibela, ukuba unomdla wokwazi okungakumbi ngayo, unokujonga kwi iinkcukacha kwikhonkco elilandelayo.