Iinkampani ezinkulu ezibandakanyekayo kuphuhliso kwinkalo yokufunda koomatshini zibonise iprojekthi ye-OpenXLA, ejolise kuphuhliso oludibeneyo lwezixhobo zokuqulunqa kunye nokuphucula imizekelo yeenkqubo zokufunda ngomatshini. Iprojekthi ithathele uphuhliso lwezixhobo ezenza ukuba kube lula 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 baye bajoyina umsebenzi odibeneyo kule projekthi.
Kulindeleke ukuba ngokudibanisa iinzame zamaqela ophando ahamba phambili kunye nabameli boluntu, kuya kwenzeka ukukhuthaza ukuphuhliswa kweenkqubo zokufunda koomatshini kunye nokusombulula ingxaki yokwahlulwa kweziseko zophuhliso kwizicwangciso ezahlukeneyo kunye nezixhobo. I-OpenXLA ikuvumela ukuba uphumeze inkxaso esebenzayo kwii-hardware ezahlukeneyo, kungakhathaliseki ukuba yeyiphi isakhelo apho imodeli yokufunda yomatshini yenziwa. I-OpenXLA kulindeleke ukuba inciphise ixesha loqeqesho lwemodeli, inyuse i-throughput, inciphise i-latency, iindleko eziphantsi ze-computing kunye nokunciphisa ixesha lokuthengisa.
I-OpenXLA iqulethe izinto ezintathu eziphambili, ikhowudi esasazwa phantsi kwelayisensi ye-Apache 2.0:
- I-XLA (i-Accelerated Linear Algebra) ngumqokeleli ovumela ukuba ukhulise iimodeli zokufunda zoomatshini ukuze usebenze ngokuphezulu kwiiplatifti ezahlukeneyo zehardware, kubandakanywa ii-GPU, ii-CPU kunye nee-accelerator ezikhethekileyo ezivela kubavelisi abohlukeneyo.
- I-StableHLO yinkcazo kunye nokuphunyezwa okusisiseko kweseti yemisebenzi ephezulu (HLO, i-High-Level Operations) yokusetyenziswa kwiimodeli zenkqubo yokufunda koomatshini. Isebenza njengomaleko phakathi kwesakhelo sokufunda koomatshini kunye nabaqulunqi abaguqula imodeli yokwenziwa kwihardware ethile. Iileyiti zokuvelisa imifuziselo kwifomathi ye-StableHLO zilungiselelwe iPyTorch, TensorFlow kunye nesikhokelo seJAX. Iseti ye-MHLO isetyenziswa njengesiseko se-StableHLO, eyandisiweyo ukuxhasa uthotho noguqulelo.
- IREE (Imeko ePhakathi yokuMelelwa kokuSebenza) ngumqambi kunye nexesha lokuqhuba eliguqula iimodeli zokufunda zoomatshini kumelo oluphakathi jikelele olusekwe kwi-MLIR (UbuMelo obuPhakathi beNqanaba eliPhakathi) kwiprojekthi ye-LLVM. Iimpawu ziquka ukubakho kokuhlanganiswa kwangaphambili (phambi kwexesha), inkxaso yokulawula ukuhamba, ukukwazi ukusebenzisa izinto eziguquguqukayo kwiimodeli, ukulungiswa kwee-CPU ezahlukeneyo kunye nee-GPU, kunye nentloko ephantsi.
Iingenelo eziphambili ze-OpenXLA toolkit:
- Phumelela ukusebenza ngokupheleleyo ngaphandle kokukhathazeka malunga nokubhala ikhowudi ethile yesixhobo. Ukubonelela ngolungelelwaniso osele lulungisiwe, kubandakanywa ukwenziwa lula kwamazwi e-algebraic, ukubekwa kwememori ngokufanelekileyo, ukucwangciswa kokwenziwa kuthathelwe ingqalelo ukunciphisa ukusetyenziswa kwememori ephezulu kunye nomthwalo ogqithisileyo.
- Ukwenziwa lula kokulinganisa kunye nokulungelelanisa izibalo. Umphuhlisi ufuna nje ukongeza izichasiselo zeseti yee-tensor ezibalulekileyo, ngokusekelwe apho umqambi angenza ngokuzenzekelayo ikhowudi ye-parallel computing.
- Qinisekisa ukuphatheka ngenkxaso yeeplatifti zehardware ezahlukeneyo, ezifana ne-AMD kunye ne-NVIDIA GPU, i-x86 kunye nee-ARM CPU, i-TPU Googl ML i-accelerators, i-AWS Trainium Inferentia IPUs, i-Graphcore kunye ne-Cerebras Wafer-Scale Engine.
- Inkxaso yokudibanisa izandiso kunye nokuphunyezwa kweempawu ezongezelelweyo, ezifana nenkxaso yokubhala i-primitives yokufunda koomatshini obunzulu usebenzisa i-CUDA, i-HIP, i-SYCL, i-Triton kunye nezinye iilwimi ze-computing efanayo. Ukunokwenzeka kokulungiswa ngesandla kweebhotile kwiimodeli.
umthombo: opennet.ru
