NgoLwesithathu (13), iRed Hat iqede inqubo yokuthenga ukuze ithole i-Neural Magic, inkampani yase-US, iphayona kusofthiwe ye-Generative Artificial Intelligence (GenAI) kanye nama-algorithms. Ubuchwepheshe be-Neural Magic kubunjiniyela bokusebenza, kanye nokuzibophezela kwayo emthonjeni ovulekile, kuhambisana nombono weRed Hat wokuletha i-AI esebenza kahle efanela izimo ezahlukene zamakhasimende kanye namacala okusetshenziswa, noma kuphi efwini elihlanganisiwe.
Nakuba isithembiso se-GenAI sibusa ingxenye enkulu yesimo sobuchwepheshe bamanje, amamodeli amakhulu olimi (ama-LLM) asekela lezi zinhlelo ayaqhubeka nokukhula. Ngenxa yalokho, ukwakha izinsizakalo ze-LLM ezithembekile nezingabizi kakhulu kudinga amandla amakhulu okusebenzisa ikhompyutha, izinsiza zamandla, kanye namakhono okusebenza akhethekile. Njengamanje, lezi zithiyo zivimbela izinhlangano eziningi ekuqapheleni izinzuzo ze-AI ephephile, elungele ukuthunyelwa, futhi eyenzelwe wena.
Ngokuthola i-Neural Magic, i-Red Hat ihlose ukubhekana nalezi zinselele ngokwenza i-GenAI ifinyeleleke kalula ezinhlanganweni eziningi ngokusebenzisa ubuchwepheshe obuvulekile be-vLLM. Yasungulwa yi-UC Berkeley, i-vLLM iyiphrojekthi yomthombo ovulekile egcinwe umphakathi yesevisi yamamodeli avulekile (indlela amamodeli e-GenAI aqonda futhi axazulula ngayo izinkinga), esekela yonke imindeni emikhulu yamamodeli, ucwaningo lokusheshisa ukuqondiswa okuthuthukisiwe, kanye nama-backend ahlukahlukene e-hardware, kufaka phakathi ama-AMD GPU, i-AWS Neuron, ama-Google TPU, i-Intel Gaudi, ama-NVIDIA GPU, kanye nama-CPU e-x86. Ubuholi be-Neural Magic kuphrojekthi ye-vLLM buhlanganiswe nephothifoliyo eqinile ye-Red Hat yobuchwepheshe be-AI yamafu ahlanganisiwe buzonikeza izinhlangano indlela evulekile yokwakha amasu e-AI ahlangabezana nezidingo zazo ezihlukile, noma ngabe idatha yazo ikhona kuphi.
KuMatt Hicks, uMongameli kanye ne-CEO yenkampani, ukuthengwa kwe-Neural Magic, kanye nokuthuthukiswa kohlelo lwe-vLLM, kuyisinyathelo sokuqala ekubekeni inkampani njengesilinganiso sobuhlakani bokwenziwa. "Sijabule kakhulu ukugcwalisa iphothifoliyo yethu ye-AI egxile efwini exubile ngokusungula i-AI entsha ye-Neural Magic, siqhubekisela phambili isifiso sethu sokuba kungabi nje 'yi-Red Hat yomthombo ovulekile,' kodwa futhi 'yi-Red Hat ye-AI,'" esho.
I-Red Hat + Umlingo Wezinzwa: Ukunika amandla ikusasa nge-AI exubile elungele amafu.
I-Neural Magic yaqala ukusebenza e-MIT ngo-2018 ngenhloso yokwakha isofthiwe yokuqonda esebenza kahle kakhulu yokufunda okujulile. Ngobuchwepheshe be-Neural Magic kanye nobuchwepheshe bobunjiniyela bokusebenza, i-Red Hat ifuna ukusheshisa umbono wayo wekusasa le-AI, eqhutshwa yiphothifoliyo yobuchwepheshe be-AI ye-Red Hat. Yakhelwe ukunqoba izinselele ze-AI yebhizinisi elikhulu, inkampani isebenzisa ukusungula izinto ezintsha ukuze ithuthukise ukufinyelela kwamandla okuguqula i-AI ngokusebenzisa:
- Amamodeli anelayisensi yomthombo ovulekile asukela kumapharamitha ayizigidi eziyi-1 kuya ku-405 billion angasebenza noma kuphi efwini elihlanganisiwe—ezikhungweni zedatha yebhizinisi, emafwini amaningi, nasemaphethelweni.
- Izici zokwenza ngokwezifiso ezivumela izinhlangano ukuthi zivumelanise kalula ama-LLM nedatha yazo yangasese futhi zisebenzise izimo ezinohlaka lokuphepha oluqinile.
- Isipiliyoni sobunjiniyela bokusebenza kokuqagela, okuholela ekusebenzeni kahle okukhulu kokusebenza kanye nengqalasizinda.
- I-ecosystem yomthombo ovulekile kanye nenethiwekhi yabalingani kanye nezakhiwo zokusekela ezinikeza amakhasimende ukukhetha okwengeziwe, kusukela kuma-LLM namathuluzi kuya ku-hardware yeseva eqinisekisiwe kanye nezakhiwo ze-chip.
Ubuholi be-vLLM Bokuthuthukisa i-Red Hat AI
I-Neural Magic izosebenzisa ubuchwepheshe bayo nolwazi ku-vLLM ukwakha isisekelo sobuchwepheshe esisezingeni lebhizinisi esivumela amakhasimende ukuthi athuthukise, asebenzise, futhi alinganise imithwalo yemisebenzi ye-LLM ezindaweni zamafu ezihlanganisiwe ngokulawula okuphelele ukukhetha ingqalasizinda, izinqubomgomo zokuphepha, kanye nomjikelezo wokuphila wamamodeli. I-Neural Magic iphinde yenze ucwaningo lokwenza ngcono amamodeli, yakhe i-LLM Compressor (umtapo wolwazi ohlangene wokwenza ngcono ama-LLM ngama-algorithms asezingeni eliphezulu kanye nokulinganisa), futhi igcina indawo yokugcina amamodeli alungiselelwe kusengaphambili alungele ukusetshenziswa nge-vLLM.
I-Red Hat AI ihlose ukusiza amakhasimende ukunciphisa izindleko ze-AI kanye nezithiyo zamakhono ngobuchwepheshe obunamandla obufana nalokhu:
- I-Red Hat Enterprise Linux AI (RHEL AI) , ipulatifomu yokwakha amamodeli okuthuthukisa, ukuhlola, nokusebenzisa umndeni we-IBM Granite wama-LLM omthombo ovulekile wezinhlelo zokusebenza zebhizinisi ekufakweni kweseva ye-Linux;
- I-Red Hat OpenShift AI iyipulatifomu ye-AI enikeza amathuluzi okuthuthukisa ngokushesha, ukuqeqesha, ukukhonza, nokuqapha amamodeli okufunda komshini ezindaweni ezisatshalaliswe ze-Kubernetes esizeni, efwini lomphakathi, noma emaphethelweni.
- I-InstructLab iphrojekthi yomphakathi evulekile eyenziwe yi-Red Hat kanye ne-IBM evumela noma ubani ukuthi akhe ikusasa le-GenAI ngokuthuthukiswa ngokubambisana kwe-Granite LLMs, enelayisensi njengomthombo ovulekile, kusetshenziswa ubuchwepheshe bokulungisa kahle be-InstructLab.
Ubuholi bezobuchwepheshe be-Neural Magic ku-vLLM buzothuthukisa ikhono le-Red Hat AI lokusekela ukuthunyelwa kwe-LLM kunoma iyiphi indawo nanoma yikuphi efwini elihlanganisiwe nge-inference stack ekulungele, elungiselelwe kakhulu, futhi evulekile.
Lokhu kuthengiselana kusangaphansi kwemvume yomthetho yase-US kanye neminye imibandela yokuvala evamile.
