{"id":107177,"date":"2023-03-10T12:48:22","date_gmt":"2023-03-10T10:48:23","guid":{"rendered":"https:\/\/prohoster.info\/?p=107177"},"modified":"2023-03-11T11:43:26","modified_gmt":"2023-03-11T09:43:26","slug":"predstavlen-openxla-instrumentarij-dlya-optimizaczii-i-kompilyaczii-modelej-mashinnogo-obucheniya","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/predstavlen-openxla-instrumentarij-dlya-optimizaczii-i-kompilyaczii-modelej-mashinnogo-obucheniya","title":{"rendered":"OpenXLA \u00ebsht\u00eb prezantuar, nj\u00eb instrument p\u00ebr optimizimin dhe kompilimimin e modeleve t\u00eb nx\u00ebnies nga makinat","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Kompani m\u00eb t\u00eb m\u00ebdha q\u00eb merren me zhvillimin n\u00eb fush\u00ebn e m\u00ebsimit me makin\u00eb kan\u00eb paraqitur projektin OpenXLA, i cili \u00ebsht\u00eb i fokusuar n\u00eb zhvillimin e p\u00ebrbashk\u00ebt t\u00eb mjeteve p\u00ebr kompilimin dhe optimizimin e modeleve p\u00ebr sistemet e m\u00ebsimit me makin\u00eb. Projekti ka p\u00ebrfshir\u00eb zhvillimin e mjeteve q\u00eb lejojn\u00eb unifikimin e kompilimit t\u00eb modeleve t\u00eb p\u00ebrgatitura n\u00eb framework-et TensorFlow, PyTorch dhe JAX, p\u00ebr nj\u00eb trajnim dhe ekzekutimin efektiv n\u00eb GPU t\u00eb ndryshme dhe akcelerator\u00eb t\u00eb specializuar. N\u00eb projektin jan\u00eb p\u00ebrfshir\u00eb kompani si Google, NVIDIA, AMD, Intel, Meta, Apple, Arm, Alibaba dhe Amazon. <\/p>\n
Pritet q\u00eb, p\u00ebrmes bashkimit t\u00eb p\u00ebrpjekjeve t\u00eb grupeve k\u00ebrkimore udh\u00ebheq\u00ebse dhe p\u00ebrfaq\u00ebsuesve t\u00eb komunitetit, t\u00eb mund\u00ebsohet stimulimi i zhvillimit t\u00eb sistemeve t\u00eb m\u00ebsimit me makin\u00eb dhe t\u00eb zgjidhen problemet e fragmentimit t\u00eb infrastruktur\u00ebs p\u00ebr framework-et dhe pajisjet e ndryshme. OpenXLA mund\u00ebson realizimin e mb\u00ebshtetjes s\u00eb efektshme p\u00ebr pajisje t\u00eb ndryshme, pa marr\u00eb parasysh se mbi \u00e7far\u00eb framework-u \u00ebsht\u00eb nd\u00ebrtuar modeli i m\u00ebsimit me makin\u00eb. Pritet q\u00eb OpenXLA t\u00eb kontribuoj\u00eb p\u00ebr t\u00eb reduktuar koh\u00ebn e trajtimit t\u00eb modeleve, p\u00ebr t\u00eb rritur kapacitetin, p\u00ebr t\u00eb zvog\u00ebluar vonesat, p\u00ebr t\u00eb indir\u00eb kostot e burimeve kompjuterike dhe p\u00ebr t\u00eb shkurtuar koh\u00ebn p\u00ebr t\u00eb sjell\u00eb produktin n\u00eb treg.
<\/center> <\/p>\n
OpenXLA p\u00ebrb\u00ebhet nga tre komponente kryesore, kodi i t\u00eb cilave shp\u00ebrndahet n\u00ebn licenc\u00ebn Apache 2.0: <\/p>\n
\n
XLA (Algebra Lineare e Acceleruar) — nj\u00eb kompilator q\u00eb lejon optimizimin e modeleve t\u00eb m\u00ebsimit t\u00eb makinerive p\u00ebr ekzekutimin me performanc\u00eb t\u00eb lart\u00eb n\u00eb platformat e ndryshme harduerike, duke p\u00ebrfshir\u00eb GPU, CPU dhe paisje t\u00eb specializuara nga prodhues t\u00eb ndrysh\u00ebm.\n
StableHLO — specifikimi dhe implementimi baz\u00eb i nj\u00eb grupi operacionesh t\u00eb nivelit t\u00eb lart\u00eb (HLO, High-Level Operations) p\u00ebr t'u p\u00ebrdorur n\u00eb modelet e sistemeve t\u00eb m\u00ebsimit t\u00eb makinerive. Vepron si nj\u00eb nd\u00ebrfaqe midis kornizave t\u00eb m\u00ebsimit t\u00eb makinerive dhe kompilator\u00ebve q\u00eb transformojn\u00eb modelin p\u00ebr ekzekutimin n\u00eb harduer t\u00eb ve\u00e7ant\u00eb. Nd\u00ebrfaqet p\u00ebr gjenerimin e modeleve n\u00eb formatin StableHLO jan\u00eb p\u00ebrgatitur p\u00ebr kornizat PyTorch, TensorFlow dhe JAX. Si baz\u00eb p\u00ebr StableHLO \u00ebsht\u00eb p\u00ebrdorur grupi MHLO, i cili \u00ebsht\u00eb zgjeruar me mb\u00ebshtetje p\u00ebr serilizimin dhe versionimin.\n
IREE (Mjedisi i Ekzekutimit t\u00eb P\u00ebrfaq\u00ebsimit t\u00eb Nd\u00ebrmjet\u00ebm) — nj\u00eb kompilator dhe runtime q\u00eb transformon modelet e m\u00ebsimit t\u00eb makinerive n\u00eb nj\u00eb p\u00ebrfaq\u00ebsim t\u00eb p\u00ebrgjithsh\u00ebm t\u00eb nd\u00ebrmjet\u00ebm, t\u00eb bazuar n\u00eb formatin MLIR (Multi-Level Intermediate Representation) nga projekti LLVM. Nga ve\u00e7orit\u00eb theksohet mund\u00ebsia e kompilimit paraprak (ahead-of-time), mb\u00ebshtetje p\u00ebr menaxhimin e rrjedh\u00ebs, mund\u00ebsia e p\u00ebrdorimit t\u00eb elementeve dinamike n\u00eb modele, optimizimi p\u00ebr CPU dhe GPU t\u00eb ndryshme, dhe kostot e ul\u00ebta operuese. <\/ul>\n
P\u00ebrfitimet kryesore t\u00eb mjeteve OpenXLA: <\/p>\n
\n
Arritja e performanc\u00ebs optimale pa nevoj\u00ebn p\u00ebr t\u00eb thelluar n\u00eb kodimin e specifik p\u00ebr pajisje t\u00eb caktuara. Ofrimi i optimizimeve t\u00eb gatshme, duke p\u00ebrfshir\u00eb thjeshtimin e shprehjeve algebrike, vendosjen efikase n\u00eb memorie, planifikimin e ekzekutimit me q\u00ebllim p\u00ebr t\u00eb zvog\u00ebluar konsumin maksimal t\u00eb memories dhe ngarkesave.\n
Thjeshtimi i shkall\u00ebzimit dhe paralelizimit t\u00eb llogaritjeve. Ajo q\u00eb duhet t\u00eb b\u00ebj\u00eb zhvilluesi \u00ebsht\u00eb t\u00eb shtoj\u00eb anotacione p\u00ebr nj\u00eb n\u00ebnmbledhje t\u00eb tenseve kritik\u00eb, mbi t\u00eb cilat kompileri mund t\u00eb gjeneroj\u00eb automatikisht kod p\u00ebr llogaritje t\u00eb paralelizuara.\n
Sigurimi i transportueshm\u00ebris\u00eb p\u00ebrmes mb\u00ebshtetjes s\u00eb platformave t\u00eb ndryshme harduerike, si GPU AMD dhe NVIDIA, CPU mbi arkitektur\u00ebn x86 dhe ARM, akcelerator\u00eb ML TPU Google, IPU AWS Trainium Inferentia, Graphcore dhe Cerebras Wafer-Scale Engine.\n
Mb\u00ebshtetje p\u00ebr lidhjen e zgjerimeve me implementimin e mund\u00ebsive shtes\u00eb, si mb\u00ebshtetje p\u00ebr kodimin e primitiveve t\u00eb thell\u00eb t\u00eb m\u00ebsimit t\u00eb makinerive duke p\u00ebrdorur CUDA, HIP, SYCL, Triton dhe gjuh\u00eb t\u00eb tjera p\u00ebr llogaritje t\u00eb paralelizuara. Mund\u00ebsia e tunimit manual t\u00eb vendeve t\u00eb ngushta n\u00eb modele. <\/ul>\n
![\"OpenXLA](\"\/wp-content\/uploads\/2023\/03\/b684d2f57def6348f6a0037c36baf717.jpg\")
<\/center> <\/p>\n