OpenXLA shine tsarin haɗin gwiwar buɗe tushen ML mai tarawa
Kwanan nan, manyan kamfanonin da suka tsunduma cikin ci gaba a fagen koyon injin da aka gabatar aikin BudeXLA, wanda aka yi nufi don haɓaka haɗin gwiwa na kayan aiki don haɗawa da haɓaka samfura don tsarin koyo na inji.
Aikin ya ɗauki nauyin haɓaka kayan aikin da ke ba da damar haɗa haɗin haɗakar samfuran da aka shirya a cikin tsarin TensorFlow, PyTorch da JAX don ingantaccen horo da aiwatarwa akan GPUs daban-daban da ƙwararrun masu haɓakawa. Kamfanoni irin su Google, NVIDIA, AMD, Intel, Meta, Apple, Arm, Alibaba da Amazon sun shiga aikin haɗin gwiwa na aikin.
Aikin OpenXLA yana samar da na'ura mai haɗawa ta ML na zamani wanda zai iya ƙima a tsakanin sarƙaƙƙiya na kayan aikin ML. Tushensa na asali shine aiki, haɓakawa, ɗaukakawa, sassauci da haɓakawa ga masu amfani. Tare da OpenXLA, muna fatan buɗe ainihin yuwuwar AI ta haɓaka haɓakawa da bayarwa.
OpenXLA yana ba masu haɓaka damar tattarawa da haɓaka samfura daga duk manyan tsare-tsaren ML don ingantaccen horo da sabis akan kayan aiki iri-iri. Masu haɓakawa da ke amfani da OpenXLA za su ga gagarumin ci gaba a lokacin horo, aiki, jinkirin sabis, da kuma lokacin kasuwa da ƙididdige farashi.
Ana fatan cewa ta hanyar shiga cikin kokarin daga cikin manyan kungiyoyin bincike da wakilan al'umma, zai yiwu a tada ci gaban tsarin koyon injin da kuma magance matsalar rarrabuwar kawuna ga ƙungiyoyi da ƙungiyoyi daban-daban.
OpenXLA yana ba da damar aiwatar da ingantaccen tallafi don kayan aiki daban-daban, ba tare da la'akari da tsarin da tsarin koyon injin ya dogara ba. Ana sa ran OpenXLA zai rage lokacin horon samfuri, haɓaka aiki, rage jinkiri, rage ƙididdiga sama da sama, da rage lokaci zuwa kasuwa.
BudeXLA ya ƙunshi manyan abubuwa guda uku, lambar wadda aka rarraba ƙarƙashin lasisin Apache 2.0:
- XLA (accelerated linear algebra) mai tarawa ne wanda ke ba ku damar haɓaka ƙirar koyon injin don aiwatar da babban aiki akan dandamalin kayan masarufi daban-daban, gami da GPUs, CPUs, da ƙwararrun masu haɓakawa daga masana'antun daban-daban.
- StableHLO ƙayyadaddun ƙayyadaddun ƙayyadaddun bayanai ne da aiwatar da saiti na Babban Ayyukan Ayyuka (HLOs) don amfani a cikin ƙirar tsarin koyo na inji. Yana aiki azaman Layer tsakanin tsarin koyo na inji da masu tarawa waɗanda ke canza ƙirar don aiki akan takamaiman kayan aiki. An shirya yadudduka don samar da samfura a cikin tsarin StableHLO don tsarin PyTorch, TensorFlow da JAX. Ana amfani da suite na MHLO azaman tushen tushen StableHLO, wanda aka haɓaka tare da tallafi don serialization da sarrafa sigar.
- IREE (Matsakaici Tsakanin Muhalli na Kisa) mai tarawa ne kuma lokaci mai aiki wanda ke canza ƙirar koyon injin zuwa wakilcin tsaka-tsakin duniya bisa tsarin MLIR (Matsakaici Multi-Level Representation) na aikin LLVM. Daga cikin fasalulluka, yuwuwar ƙaddamarwa (kafin lokaci), tallafi don sarrafa kwarara, ikon yin amfani da abubuwa masu ƙarfi a cikin ƙira, haɓakawa don CPUs daban-daban da GPUs, da ƙarancin sama da ƙasa ana haskaka su.
Game da manyan fa'idodin OpenXLA, an ambaci hakan an sami kyakkyawan aiki ba tare da zurfafa cikin lambar rubutu ba takamaiman na'urar, ban da samar da abubuwan ingantawa a waje, ciki har da sauƙaƙan maganganun algebra, ingantaccen rarraba ƙwaƙwalwar ajiya, tsara tsarin aiwatarwa, la'akari da raguwar matsakaicin yawan amfanin ƙwaƙwalwar ajiya da wuce gona da iri.
Wata fa'ida ita ce sauƙaƙan ƙira da daidaitawa na ƙididdiga. Ya isa ga mai haɓakawa don ƙara bayanan bayanai don juzu'in mahimmancin tenors, akan abin da mai tarawa zai iya samar da lamba ta atomatik don lissafin layi ɗaya.
An kuma haskaka cewa ana ba da damar ɗauka tare da goyan bayan dandamali na kayan aiki da yawa, irin su AMD da NVIDIA GPUs, x86 da ARM CPUs, Google TPU ML Accelerators, AWS Trainium Inferentia IPUs, Graphcore, da Wafer-Scale Engine Cerebras.
Taimako don haɗa haɓakawa tare da aiwatar da ƙarin ayyuka, a matsayin tallafi don rubuta zurfin koyan na'ura ta amfani da CUDA, HIP, SYCL, Triton da sauran yarukan don lissafin layi ɗaya, da kuma yuwuwar daidaitawa da hannu na kwalabe a cikin samfura.
A ƙarshe, idan kuna da sha'awar sanin ƙarin abubuwa game da shi, kuna iya tuntuɓar cikakkun bayanai a cikin mahaɗin mai zuwa.