I-OpenXLA, iphrojekthi yomthombo ovulekile yokusheshisa nokwenza lula ukufunda komshini

Muva nje izinkampani ezinkulu ezibambe iqhaza ekuthuthukisweni emkhakheni wokufunda ngomshini owethulwe iphrojekthi I-OpenXLA, okuhloswe ukuthuthukiswa okuhlangene kwamathuluzi ukuhlanganisa nokuthuthukisa amamodeli wezinhlelo zokufunda zomshini.

Le phrojekthi ibambe iqhaza ekuthuthukisweni kwamathuluzi avumela ukuhlanganisa ukuhlanganiswa kwamamodeli alungiswe kuzinhlaka ze-TensorFlow, PyTorch kanye ne-JAX zokuqeqeshwa okuphumelelayo nokusebenza kuma-GPU ahlukene kanye nama-accelerator akhethekile. Izinkampani ezifana ne-Google, NVIDIA, AMD, Intel, Meta, Apple, Arm, Alibaba kanye ne-Amazon zijoyine umsebenzi ohlanganyelwe wephrojekthi.

Iphrojekthi ye-OpenXLA ihlinzeka ngenhlanganisela yesimanje ye-ML ekwazi ukukala phakathi kobunzima bengqalasizinda ye-ML. Izinsika zayo eziyisisekelo ukusebenza, ukunwebeka, ukuphatheka, ukuguquguquka nokunwebeka kubasebenzisi. Nge-OpenXLA, sifisa ukuvula amandla angempela e-AI ngokusheshisa ukuthuthukiswa nokulethwa kwayo.

I-OpenXLA inika amandla abathuthukisi ukuthi bahlanganise futhi balungiselele amamodeli avela kuzo zonke izinhlaka eziholayo ze-ML ukuze uthole ukuqeqeshwa okuphumelelayo nokuseviswa ezinhlobonhlobo zehadiwe. Onjiniyela abasebenzisa i-OpenXLA bazobona ukuthuthuka okubalulekile esikhathini sokuqeqeshwa, ukusebenza, ukubambezeleka kwesevisi, futhi ekugcineni nesikhathi sokumaketha nokubala izindleko.

Kwethenjwa ukuthi ngokujoyina imizamo emaqenjini amakhulu ocwaningo kanye nabamele umphakathi, kuzokwazi ukugqugquzela ukuthuthukiswa kwezinhlelo zokufunda zomshini kanye nokuxazulula inkinga yokuhlukaniswa kwengqalasizinda yezinhlaka namathimba ahlukahlukene.

I-OpenXLA ivumela ukusebenzisa ukusekelwa okusebenzayo kwehadiwe ehlukahlukene, kungakhathaliseki uhlaka lapho imodeli yokufunda yomshini isekelwe khona. I-OpenXLA kulindeleke ukuthi yehlise isikhathi sokuqeqeshwa okuyimodeli, ithuthukise ukusebenza, inciphise ukubambezeleka, inciphise i-computing overhead, futhi yehlise isikhathi sokumaketha.

I-OpenXLA iqukethe izingxenye ezintathu ezibalulekile, ikhodi esatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0:

1. I-XLA (i-algebra yomugqa esheshisiwe) iyinhlanganisela ekuvumela ukuthi ulungiselele amamodeli okufunda omshini ukuze asebenze kahle ezinkundleni zehadiwe ezahlukahlukene, okuhlanganisa ama-GPU, ama-CPU, nama-accelerator akhethekile avela kubakhiqizi abahlukahlukene.
2. I-StableHLO iwukucaciswa okuyisisekelo nokusebenzisa isethi Yemisebenzi Yezinga Eliphezulu (ama-HLO) ukuze isetshenziswe kumamodeli esistimu yokufunda komshini. Isebenza njengesendlalelo phakathi kwezinhlaka zokufunda zomshini nabahlanganisi abaguqula imodeli ukuze isebenze kuzingxenyekazi zekhompuyutha ezithile. Izendlalelo zilungiselelwa ukukhiqiza amamodeli ngefomethi ye-StableHLO yezinhlaka ze-PyTorch, TensorFlow kanye ne-JAX. I-MHLO suite isetshenziswa njengesisekelo se-StableHLO, esinwetshwa ngokusekelwa kochungechunge nokulawulwa kwenguqulo.
3. I-IREE (I-Intermediate Representation Execution Environment) ihlanganisa nesikhathi sokuqalisa esiguqula amamodeli okufunda omshini abe ukumelwa okumaphakathi kwendawo yonke okusekelwe kufomethi ye-MLIR (I-Intermediate Multi-Level Representation) yephrojekthi ye-LLVM. Ezicini, ithuba lokuhlanganisa kusengaphambili (ngaphambi kwesikhathi), ukusekelwa kokulawula ukugeleza, ikhono lokusebenzisa izakhi eziguqukayo kumamodeli, ukulungiselelwa ngokugcwele kwama-CPU nama-GPU ahlukene, kanye nokuphezulu okuphansi kuyagqanyiswa.

Mayelana nezinzuzo eziyinhloko ze-OpenXLA, kushiwo lokho ukusebenza okusezingeni eliphezulu kuzuziwe ngaphandle kokuzama ukubhala ikhodi okuqondene nedivayisi, ngaphezu kwalokho hlinzeka ngokulungiselelwa okungaphandle kwebhokisi, okuhlanganisa ukwenziwa lula kwezinkulumo ze-algebraic, ukwaba inkumbulo okuphumelelayo, ukuhlela ukwenza, kucatshangelwa ukuncishiswa kokusetshenziswa kwememori okuphezulu kanye nokweqa.

Enye inzuzo yi- ukwenza kube lula ukukala kanye nokuhambisana kwezibalo. Kwanele ukuthi umthuthukisi angeze izichasiselo zesethi engaphansi yama-tensors abalulekile, ngesisekelo lapho umdidiyeli angakwazi ukukhiqiza ngokuzenzakalelayo ikhodi ye-parallel computing.

Kubuye kuqhakanjiswe lokho ukuphatheka kunikezwa ngosekelo lwezingxenyekazi zehadiwe eziningi, njenge-AMD ne-NVIDIA GPUs, i-x86 ne-ARM CPUs, i-Google TPU ML Accelerators, i-AWS Trainium Inferentia IPUs, i-Graphcore, ne-Wafer-Scale Engine Cerebras.

Ukusekelwa kokuxhuma izandiso ngokusetshenziswa kwemisebenzi eyengeziwe, njengokwesekwa kokubhala ama-primitives okufunda komshini ojulile usebenzisa i-CUDA, i-HIP, i-SYCL, i-Triton nezinye izilimi ze-computing efanayo, kanye kungenzeka ukulungiswa ngesandla kwamabhodlela kumamodeli.

Ekugcineni, uma unentshisekelo yokwazi okwengeziwe ngakho, ungaxhumana ne- imininingwane kusixhumanisi esilandelayo.