{"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\/sk\/blog\/news\/predstavlen-openxla-instrumentarij-dlya-optimizaczii-i-kompilyaczii-modelej-mashinnogo-obucheniya","title":{"rendered":"Predstaven\u00fd OpenXLA, n\u00e1stroj na optimaliz\u00e1ciu a kompil\u00e1ciu modelov strojov\u00e9ho u\u010denia","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Najv\u00e4\u010d\u0161ie spolo\u010dnosti zaoberaj\u00face sa v\u00fdvojom v oblasti strojov\u00e9ho u\u010denia predstavili projekt OpenXLA, ktor\u00fd je zameran\u00fd na spolo\u010dn\u00fd rozvoj n\u00e1strojov na kompil\u00e1ciu a optimaliz\u00e1ciu modelov pre syst\u00e9my strojov\u00e9ho u\u010denia. Do projektu pre\u0161li n\u00e1stroje, ktor\u00e9 umo\u017e\u0148uj\u00fa unifikovan\u00fa kompil\u00e1ciu modelov pripraven\u00fdch vo frameworkoch TensorFlow, PyTorch a JAX, pre efekt\u00edvne u\u010denie a vykon\u00e1vanie na r\u00f4znych GPU a \u0161pecializovan\u00fdch akceler\u00e1toroch. Do spolupr\u00e1ce na projekte sa zapojili tak\u00e9 spolo\u010dnosti ako Google, NVIDIA, AMD, Intel, Meta, Apple, Arm, Alibaba a Amazon. <\/p>\n
O\u010dak\u00e1va sa, \u017ee v\u010faka zl\u00fa\u010deniu s\u00edl ved\u00facich v\u00fdskumn\u00fdch t\u00edmov a z\u00e1stupcov komunity sa podar\u00ed stimulova\u0165 rozvoj syst\u00e9mov strojov\u00e9ho u\u010denia a vyrie\u0161i\u0165 probl\u00e9m s fragment\u00e1ciou infra\u0161trukt\u00fary pre r\u00f4zne frameworky a vybavenie. OpenXLA umo\u017e\u0148uje implementova\u0165 efekt\u00edvnu podporu r\u00f4zneho vybavenia, bez oh\u013eadu na to, na akom frameworku bola model strojov\u00e9ho u\u010denia vytvoren\u00e1. O\u010dak\u00e1va sa, \u017ee v\u010faka OpenXLA sa podar\u00ed skr\u00e1ti\u0165 \u010das u\u010denia modelov, zv\u00fd\u0161i\u0165 priepustnos\u0165, skr\u00e1ti\u0165 oneskorenia, zn\u00ed\u017ei\u0165 n\u00e1klady na v\u00fdpo\u010dtov\u00e9 zdroje a skr\u00e1ti\u0165 \u010das uvedenia produktu na trh.
<\/center> <\/p>\n
OpenXLA tvoria tri hlavn\u00e9 komponenty, ktor\u00fdch k\u00f3d je \u0161\u00edren\u00fd pod licenciou Apache 2.0: <\/p>\n
\n
XLA (Accelerated Linear Algebra) \u2014 kompil\u00e1tor, ktor\u00fd umo\u017e\u0148uje optimalizova\u0165 modely strojov\u00e9ho u\u010denia pre v\u00fdkonn\u00e9 vykon\u00e1vanie na r\u00f4znych hardv\u00e9rov\u00fdch platform\u00e1ch, vr\u00e1tane GPU, CPU a \u0161pecializovan\u00fdch akceler\u00e1torov od r\u00f4znych v\u00fdrobcov.\n
StableHLO \u2014 \u0161pecifik\u00e1cia a z\u00e1kladn\u00e1 implement\u00e1cia s\u00faboru vysoko\u00farov\u0148ov\u00fdch oper\u00e1ci\u00ed (HLO, High-Level Operations) na pou\u017eitie v modeloch syst\u00e9mov strojov\u00e9ho u\u010denia. Sl\u00fa\u017ei ako vrstva medzi r\u00e1mcami strojov\u00e9ho u\u010denia a kompil\u00e1tormi, ktor\u00e9 transformuj\u00fa model na vykon\u00e1vanie na konkr\u00e9tnom hardv\u00e9ri. Vrstvy na generovanie modelov vo form\u00e1te StableHLO s\u00fa pripraven\u00e9 pre r\u00e1mce PyTorch, TensorFlow a JAX. Ako z\u00e1klad pre StableHLO bol pou\u017eit\u00fd s\u00fabor MHLO, ktor\u00fd bol roz\u0161\u00edren\u00fd o podporu serializ\u00e1cie a verzionovania.\n
IREE (Intermediate Representation Execution Environment) \u2014 kompil\u00e1tor a runtime, ktor\u00fd transformuje modely strojov\u00e9ho u\u010denia do univerz\u00e1lneho medzi\u00farov\u0148ov\u00e9ho zast\u00fapenia zalo\u017een\u00e9ho na form\u00e1te MLIR (Multi-Level Intermediate Representation) od projektu LLVM. Medzi vlastnosti patr\u00ed mo\u017enos\u0165 predkompil\u00e1cie (ahead-of-time), podpora riadenia toku, mo\u017enos\u0165 pou\u017eitia dynamick\u00fdch prvkov v modeloch a optimaliz\u00e1cia pre r\u00f4zne CPU a GPU, n\u00edzke prev\u00e1dzkov\u00e9 n\u00e1klady. <\/ul>\n
Dosiahnutie optim\u00e1lneho v\u00fdkonu bez nutnosti hlb\u0161ieho zap\u00e1jania do p\u00edsania k\u00f3du \u0161pecifick\u00e9ho pre konkr\u00e9tne zariadenia. Poskytnutie hotov\u00fdch optimaliz\u00e1ci\u00ed, vr\u00e1tane zjednodu\u0161enia alge\u00adbraick\u00fdch v\u00fdrazov, efekt\u00edvneho umiestnenia v pam\u00e4ti a pl\u00e1novania vykon\u00e1vania s oh\u013eadom na zn\u00ed\u017eenie \u0161pi\u010dkovej spotreby pam\u00e4te a pre\u0165a\u017eenia.\n
Zjednodu\u0161enie \u0161k\u00e1lovania a paraleliz\u00e1cie v\u00fdpo\u010dtov. V\u00fdvoj\u00e1rovi sta\u010d\u00ed prida\u0165 anot\u00e1cie pre podmno\u017einu kritick\u00fdch tenzorov, na z\u00e1klade ktor\u00fdch kompil\u00e1tor m\u00f4\u017ee automaticky vygenerova\u0165 k\u00f3d pre paraleln\u00e9 v\u00fdpo\u010dty.\n
Zabezpe\u010denie prenosite\u013enosti prostredn\u00edctvom podpory r\u00f4znych hardv\u00e9rov\u00fdch platforiem, ako s\u00fa GPU AMD a NVIDIA, CPU zalo\u017een\u00e9 na architekt\u00fare x86 a ARM, ML-ur\u00fdch\u013eova\u010de TPU Google, IPU AWS Trainium Inferentia, Graphcore a Cerebras Wafer-Scale Engine.\n
Podpora pripojenia roz\u0161\u00edren\u00ed s implementovan\u00edm \u010fal\u0161\u00edch funkci\u00ed, ako je podpora p\u00edsania primit\u00edvov hlbok\u00e9ho strojov\u00e9ho u\u010denia pomocou CUDA, HIP, SYCL, Triton a in\u00fdch jazykov pre paraleln\u00e9 v\u00fdpo\u010dty. Mo\u017enos\u0165 manu\u00e1lneho ladenia \u00fazkych miest v modeloch. <\/ul>\n
![\"Predstaven\u00fd](\"\/wp-content\/uploads\/2023\/03\/b684d2f57def6348f6a0037c36baf717.jpg\")
<\/center> <\/p>\n