Isofthiwe ye-Exo AI: iqoqo lasekhaya lamamodeli ezilimi ezinkulu

Ukuqaliswa kwendawo kwamamodeli ezilimi ezinkulu (LLM) ivame ukuhlotshaniswa nemishini emikhulu, amakhadi ehluzo abizayo kakhulu, kanye nokusetshenziswa kwamandla okungenabungane. Kodwa-ke, ezinyangeni zamuva nje kuvele enye indlela edonsela ukunaka okukhulu: I-Exo, isofthiwe eguqula amadivayisi amaningi asekhaya abe yiqoqo le-AI ikwazi ukusebenzisa amamodeli adinga kakhulu ngaphandle kokuthembela efwini.

Nge-Exo, Umqondo uwukuhlanganisa amandla amakhompyutha, amalaptopomakhalekhukhwini namanye amadivayisi axhunyiwe Ukuze, ngokuhlanganisa izinsiza zabo, bakwazi ukusebenzisa ama-LLM esivame ukuwabona kuphela ezikhungweni zedatha zobungcweti. Lokhu kuvula umnyango wokusetha uhlobo "lwesikhungo sedatha esincane" ekhaya, esilungele amaphrojekthi afana nabasizi bendawo njengoMsizi Wasekhaya, ama-SME afuna ukuzama i-AI ngaphandle kokuchitha imali eningi, noma abathuthukisi abanelukuluku abafuna ukwenza okungaphezu kokuhlolwa kwamafu okuvamile.

Iyini i-Exo futhi ihlose ukuyixazulula kanjani?

I-Exo iyisofthiwe yomthombo ovulekile yokuhlolainakekelwa yi-Exo Labs, enhloso yayo eyinhloko ukukusiza ukuthi wakhe eyakho iqembu le ukuhlakanipha okufakelwayo ekhaya kusetshenziswa amadivayisi avamile. Esikhundleni sokuthembela ku-GPU eyodwa ene-VRAM eningi, i-Exo isabalalisa imodeli kumadivayisi amaningana axhunywe kunethiwekhi efanayo.

Inhloso ka-Exo ilula kodwa inamandla kakhulu.: amadivayisi akho avamile—i-MacBook, i- iPhoneI-iPad, i-Linux PC, noma amakhompyutha athile ebhodi elilodwa (ama-SBC) angasebenzisana ukuze asebenzise imodeli enkulu ye-AI. Ngayinye inikela ngamandla enkumbulo kanye nekhompyutha, futhi isofthiwe ihlela konke ukuze kusebenze sengathi umshini owodwa, omkhulu.

I-Exo yazalwa njengempendulo yomkhawulo ovamile kakhuluAkuwona wonke umuntu okwazi ukufinyelela i-GPU ephezulu njenge- I-NVIDIA I-RTX 3090 ene-VRAM engu-24 GB kanye nama-gigabytes ayishumi e-RAM. Ukusebenzisa imodeli "ephansi" njenge-Llama 3.1 70B esebenza kahle endaweni kudinga ihadiwe engaphezu kwesabelomali sabathuthukisi abaningi, abakhiqizi, noma amabhizinisi amancane.

Esikhundleni sokukuphoqa ukuthi utshale imali emshinini omkhulu noma uthembele kuphela efwini, I-Exo iphakamisa ukuhlanganisa inkumbulo etholakalayo yamadivayisi amaningana alula ukuze imodeli ephelele ifaneleke ngaphakathi kweqoqo lonke. Lokhu kuvumela kakhulu ukufinyelela ekuhlolweni nge-AI ethuthukisiwe, okuthakazelisa kakhulu izinkampani ezintsha kanye nama-SME.

Indlela i-Exo esebenza ngayo ezingeni lobuchwepheshe

Inhliziyo ye-Exo isekhonweni layo lokusabalalisa amamodeli e-LLM kuwo wonke amadivayisi atholakalayo kunethiwekhi yasekhaya noma yehhovisi. Esikhundleni sokulayisha yonke imodeli emshinini owodwa, i-Exo iyayihlukanisa bese iyisabalalisa, isebenzisa inkumbulo ehlanganisiwe kanye namandla ekhompyutha e-node ngayinye.

Ukwenza, I-Exo isebenzisa indlela yememori eyabiwe esabalalisiwengokushintsha ingxenye yemodeli ebekwe kudivayisi ngayinye ngokwenani lememori (i-RAM kanye, ku-macOS, i-GPU VRAM) etholakalayo. Ngakho-ke, imodeli evame ukudinga i-GPU enamandla kakhulu ingasebenza kuma-laptop, kuma-desktop, nakwamanye amadivayisi, uma nje imemori yeqoqo iyonke yanele.

Inqubo iqala ngokufakwa kwe-Exo kumadivayisi ahambisanayo.Ezimweni eziningi, kudinga ukuba Python 3.12 noma ngaphezulu, ngaphezu kwamanye amathuluzi afana nomphathi wokuxhomekeka kanye nabahlanganisi. Uma isevisi isisebenza, amadivayisi asebenzisa i-Exo bathola omunye nomunye ngokuzenzakalelayo ngaphakathi kwenethiwekhi efanayo, ngaphandle kokuthi ulungiselele ama-IP noma ama-port ngesandla.

Ukuxhumana kwangaphakathi phakathi kwama-node kusekelwe ekuxhumeni asinge-to-peerLokhu kusiza ukwakheka lapho umshini ngamunye unikela ngalokho ongakwenza ngaphandle kokuthembela kuseva eyodwa eqinile ephakathi. Ngokusekelwe kulolu hlu lwezici, i-Exo yenza ukuhlaziywa kwezinsiza kanye nokubambezeleka ukuze inqume ukuthi ingahlukanisa kanjani imodeli nokuthi kufanele ibeke kuphi ingxenye ngayinye.

Esinye sezici ezibalulekile ukwahlukaniswa okunamandla kwamamodeliI-Exo ithola umbono wesikhathi sangempela we-topology yeqembu, izinsiza zedivayisi ngayinye (inkumbulo, i-CPU, i-GPU), kanye nekhwalithi yezixhumanisi zenethiwekhi (ukubambezeleka, umkhawulokudonsa phakathi kwama-node). Ngalolu lwazi, inquma ngokuzenzakalelayo ukuthi izohlukanisa kanjani imodeli ukuze ilinganise umthwalo futhi inciphise izithiyo.

Izici zobuchwepheshe eziyinhloko ze-Exo

Iphrojekthi ye-Exo ifaka izici eziningana ezenzelwe ukusizakala kakhulu kuyo. ihadiwe etholakalayo futhi yenza impilo ibe lula kumsebenzisi. Phakathi kwezinto eziphawuleka kakhulu yilezi:

Ukutholwa kwedivayisi okuzenzakalelayoAmadivayisi asebenzisa i-Exo angahlangana ngaphandle kokuthi uthinte ukucushwa kwenethiwekhi. Lokhu kugwema iphupho elibi elivamile lokubhekana nama-IP angaguquki, ama-port asetshenziswa ngesandla, noma izikripthi ezenziwe ngokwezifiso zama-node ukuze abonelane.

Ukusekelwa kwe-RDMA ngaphezu kwe-ThunderboltElinye lamaphuzu agqamile kakhulu ukuthi i-Exo iza ne Ukuhambisana kosuku lokuqala ne-RDMA (Remote Direct Memory Access) ngaphezu kwe-Thunderbolt 5Le ndlela ivumela ukubambezeleka phakathi kwamadivayisi ukuthi kuncishiswe kufike ku-99% uma kuqhathaniswa nezindlela zokuxhumana zendabuko, okubaluleke kakhulu lapho kusatshalaliswa izingxenye zemodeli efanayo phakathi kwemishini eminingana.

Ukulinganisa kwe-tensoryI-Exo isebenzisa amasu e- i-tensor parallelismNgamanye amazwi, ihlukanisa ama-tensor emodeli kumadivayisi amaningi. Ngenxa yalokhu, kwezinye izilungiselelo, kuye kwabikwa ukuthuthukiswa kwesivinini okufika izikhathi ezingu-1,8 kusetshenziswa amadivayisi amabili kanye nokufika izikhathi ezingu-3,2 namane, uma kuqhathaniswa nokusebenzisa imodeli kudivayisi eyodwa enezinsiza ezifanayo.

I-topology eqondayo kanye nokuzifanisaUhlelo lubala ngokuzenzakalelayo indlela engcono kakhulu yokuhlukanisa imodeli bese luyisabalalisa kulo lonke iqembu, kucatshangelwa amandla e-node ngayinye kanye nekhwalithi yokuxhumana phakathi kwayo. Lesi sici se "ukuhambisana okuzenzakalelayo okuqaphela i-topology" Kunciphisa isidingo sokuba umsebenzisi alungise konke ngesandla.

Ukuhlanganiswa kwe-MLX kanye nokusekelwa kwe-GPU ku-macOSKu-macOS, i-Exo isebenzisa I-MLX njenge-backend yokuphetha kanye ne-MLX Distributed yokuxhumana okusatshalaliswe. Lokhu kuyivumela ukuthi isebenzise ama-GPU ahlanganiswe nama-chip e-Apple Silicon, okuthakazelisa kakhulu kumadivayisi afana ne- Mac I-Studio noma i-MacBook Pro yesizukulwane sakamuva.

Amacala okusetshenziswa okusebenzayo kanye nokusebenza kuma-Mac

I-Exo ihlolwe ezimweni ezinzima kakhulu Ngokuhlanganisa ama-Mac amaningana aphezulu, sisebenzisa kokubili i-tensor parallelism kanye ne-RDMA ngaphezu kwe-Thunderbolt 5. Kulezi zisetho, kuye kwasetshenziswa amamodeli amakhulu esivame ukuwabona kuphela ezindaweni zesikhungo sedatha.

Isibonelo esisodwa okuxoxwe ngaso emphakathini ukusetshenziswa i-Mac Studio ezine ene-chip ye-M3 Ultra ixhunywe ne-Thunderbolt 5 kanye ne-RDMA. Ndawonye zinezela cishe ama-TB angu-15 e-VRAM asebenzayo ukuze zivumelane namamodeli amakhulu ngendlela emangalisayo.

Phakathi kwamamodeli asetshenziswe kulolu hlobo lokusetha, okulandelayo kuyaphawuleka: I-Qwen3-235B ngamabhithi angu-8, I-DeepSeek v3.1 671B nayo ilinganiswe ngamabhithi angu-8 noma ngisho Ukucabanga kukaKimi K2 ngama-bits angu-4Kuzo zonke lezi zimo, i-Mac ngayinye ibamba ingxenye yemodeli futhi ukuxhumana phakathi kwama-node kwenziwa ngcono kusetshenziswa i-RDMA.

Lezi zivivinyo zibhalwe phansi ngabadali abanjengoJeff Geerlingokubonise ukuthi kungenzeka kanjani ukusetha ukucushwa kwekhaya ngenani elikhulu lememori yevidiyo esebenzayo. Kulesi simo, i-Exo isebenza njengesendlalelo sokuhlelwa esivumela imodeli ukuthi isebenze ngokuhambisanayo naphezu kokusatshalaliswa emishinini eminingana.

Ngale kwalezi zilungiselelo eziqine kakhulu, Isimiso esifanayo singasetshenziswa ekusetshenzisweni okunesizotha kakhuluIsibonelo, ukondla imodeli enkulu yolimi esebenza njengobuchopho bohlelo oluthuthukisiwe lokuzenzakalela ekhaya, umsizi wengxoxo yangasese, noma amathuluzi okuhlaziya umbhalo ngaphakathi kwebhizinisi elincane.

Ukusebenzisa i-Exo nge-Home Assistant kanye nokwenza izinto ngokuzenzakalela kwekhaya okuthuthukisiwe

Umphakathi we-Home Assistant usuvele ukhuluma nge-Exo njengenketho yokusebenzisa ama-LLM amakhulu endaweni asebenza njengezinjini zokucabanga zokwenza izinto ngokuzenzakalela ekhaya. Nakuba abantu abaningi bengakayizami ngqo okwamanje, le phrojekthi idonsele ukunaka ngoba ithembisa ukuvumela amamodeli asindayo kakhulu kwihadiwe encane kakhuluuma nje amaqembu anele eqoqwe ndawonye.

Umqondo uthi, esikhundleni sokuthembela ku-GPU enamandla noma efwiniUngamisa i-cluster enamakhompyutha amaningi, noma ngabe angenayo i-GPU ezinikele. I-Exo izosingatha ukusabalalisa imodeli kuwo wonke, okuvumela uMsizi Wekhaya ukuthi axhumane ne-LLM yendawo ukuze enze izinqumo eziyinkimbinkimbi noma anikeze izimpendulo zemvelo.

Le ndlela ikhanga kakhulu labo ababeka ubumfihlo phambili.Njengoba idatha yasekhaya ingadingi ukuphuma kunethiwekhi yendawo, konke ukuqagela kanye nokucubungula ulimi lwemvelo kwenziwa ngaphakathi kweqoqo olimisile, ngaphandle kokuthembela kumaseva angaphandle.

Le phrojekthi ithole ukunakwa kumabhulogi obuchwepheshe afana ne-CNX Software, lapho Kugqanyisiwe ukuthi i-Exo ingasebenza ngisho nakumaqoqo ekhompyutha, ama-smartphone, noma amabhodi okuthuthukisa.Kodwa-ke, ekusebenzeni kwamanje, ukwesekwa okucolisisiwe kakhulu kutholakala ku-macOS naku-Linux, kanti eminye imikhawulo isekhona ku- iOS.

Kwabaningi abathanda ukwenza izinto ngokuzenzakalela ekhaya, inhlanganisela ye-Exo + Home Assistant Kungaba yinto engekho ukuba nomsizi wengxoxo osezingeni eliphezulu, ohlala evuliwe, futhi okhuluma nabantu bendawo ngokuphelele, ngaphandle kokukhokhela amathokheni efu noma ukuncika kubantu besithathu.

Amamodeli e-AI ahambisanayo kanye nemikhawulo enamamodeli azimele

I-Exo iyahambisana nenani elihle lamamodeli e-AI athandwayoikakhulukazi lezo eziyingxenye ye-ecosystem evulekile futhi ezingalandwa futhi zisetshenziswe endaweni. Lokhu kufaka phakathi imindeni efana ne-LLaMA, i-Mistral, i-LLaVA, i-Qwen, ne-DeepSeek, phakathi kwabanye.

Esimweni esithile I-Llama 3 ye-MetaI-Exo isebenza kahle kakhulu ngalezi zakhiwo, okuvumela ukuhlukahluka kobukhulu obuhlukene kuye ngezinsizakusebenza zeqoqo. Lokhu kwenza kube lula kokubili abahloli ngabanye kanye namabhizinisi amancane ukuba nemodeli yolimi yesimanje ngaphandle kokuthola izindleko zelayisensi noma ukuthembela kuma-API ezinkampani zangaphandle.

Enye inzuzo ethokozisayo ukuthi kungenzeka ukuthi kudalulwe i-API ehambisana ne- I-ChatGPTNgokulungiswa okuthile, i-Exo ivumela amamodeli asebenza kuqoqo lakho ukuthi amukele izicelo ngefomethi efana ne-API ye I-OpenAIokwenza kube lula kakhulu ukuhlanganiswa nezinhlelo zokusebenza ezisebenza kakade naleyo nqubo.

Nokho, Kunomkhawulo ocacile uma sikhuluma ngamamodeli azimele njenge-GPT-4.I-GPT-4 iyimodeli evaliwe, ephethwe kwingqalasizinda ye-OpenAI futhi ayitholakali ukuthi ilandwe. Lokhu kusho ukuthi, ngokwencazelo, ayikwazi ukusebenza ku-Exo noma kunoma iyiphi enye indawo yasendaweni ngaphandle kwe-OpenAI.

Ngakho-ke, uma usebenza ne-Exo Kumelwe unamathele kumamodeli omthombo ovulekile noma lawo atholakalayo ukuze asetshenziswe endaweniEziningi zezinhlelo zokusebenza ezisebenzayo (ama-chatbot, abasizi, ukuhlaziywa kombhalo, ukucabanga okuyisisekelo nokuthuthukile) sekuvele kunekhathalogi enamandla ngokwanele yamamodeli omthombo ovulekile, ikakhulukazi ngokuvela kwamuva kwamamodeli afana ne-Llama 3.x, i-Qwen noma i-DeepSeek.

Izinzuzo eziyinhloko zokusebenzisa i-Exo cluster

Inzuzo enkulu ye-Exo ukuthi ivumela ukwehla okukhulu kokutshalwa kwezimali kwehadiwe.Esikhundleni sokuthenga i-workstation ene-GPU enkulu, ungasebenzisa kabusha amakhompyutha namadivayisi osuvele unawo ekhaya noma ehhovisi. Abasebenzisi abaningi bangasetha i-cluster ngokusebenzisa ama-Mac abo, i-PC endala, kanye neminye imishini engasetshenziswa kahle.

Le ndlela yenza kube lula ukufinyelela ku-AI ethuthukisiweLokhu kuvumela amabhizinisi amasha kanye nama-SME anesabelomali esincane ukuthi ancintisane kuligi eyayigcinelwe izinkampani ezinezinsizakusebenza eziningi ngaphambilini. Akuwona wonke amaphrojekthi avumela ukuqasha ama-GPU abizayo efwini, futhi yilapho iqembu lendawo linengqondo khona.

Futhi, I-Exo iguquguquka kakhulu ngokwezinhlobo zamadivayisiNakuba ukwesekwa okuvuthiwe kakhulu okwamanje kukuma-macOS (kusetshenziswa ama-GPU) kanye ne-Linux (okwamanje kusetshenziswa ama-CPU), iphrojekthi ihlose ukuhlanganisa nezinye izinhlobo zehadiwe kanye nokuthuthukisa ukuhambisana ngamaqhinga. Leli khono lokuxuba amaphrofayili ahlukene omshini likuvumela ukuthi uklame i-cluster ngokwezidingo zakho ezithile.

Enye inzuzo ebalulekile ukuthi, ngenxa yokwakheka kwayo okusabalele, Iqoqo lingalinganiswa kalulaUma udinga ukusebenza okwengeziwe nganoma yisiphi isikhathi, mane nje wengeze enye idivayisi esebenzisa i-Exo. I-node ngayinye entsha izotholakala ngokuzenzakalelayo futhi izinsiza zayo zengezwe kusamba, ngaphandle kokuthi kudingeke ukuthi uklame kabusha uhlelo kusukela ekuqaleni.

Ekugcineni, ukuqwashisa ngokuhlukanisa okunamandla kanye nokuqwashisa nge-topology Baqinisekisa ukuthi ngisho namadivayisi anehadiwe encane angafaka isandla entweni ewusizo. Nakuba i-laptop endala ingeke ikwazi ukusebenzisa lonke uhlelo ngokwayo, ingakwazi ukusingatha omunye womsebenzi noma yenze ezinye zezibalo, inikele ekusebenzeni okuphelele.

Imikhawulo yamanje namaphuzu okufanele acatshangelwe

Akuzona zonke izinto eziphelele, futhi i-Exo isalokhu iyisofthiwe yokuhlola ngokusobala.Lokhu kusho ukuthi amaphutha, ukuziphatha okungazinzile, noma izinguquko ezivame ukwenzeka endleleni uhlelo oluhlelwe futhi olusebenza ngayo kungase kwenzeke. Akusona isisombululo sokuthi "setha bese ukhohlwa" ezingeni lomkhiqizo wezohwebo ovuthiwe.

Kumadivayisi e-iOS afana ne-iPhone ne-iPad, Ukusekelwa kusalinganiselwe Futhi kudinga ukuthuthukiswa okwengeziwe ukuze kusetshenziswe ngempela emphakathini jikelele. Nakuba lo msebenzi uhlose ukusebenzisa la madivayisi futhi, okwamanje iningi lolwazi olubushelelezi lutholakala kumakhompyutha.

Esinye isici esibalulekile ukuthi Ukusebenza kahle kweqoqo kuzoncika kakhulu kunethiwekhi yasekhaya.Uma unenethiwekhi ye-Wi-Fi eminyene, ama-router aphelelwe yisikhathi, noma ukuxhumana okuhamba kancane, ukusabalalisa imodeli enkulu kuma-node amaningi kungaba yisithiyo esikhulu. Ezindaweni ezibucayi, cishe kuzomele usebenzise ukuxhumana okunezintambo okusezingeni eliphezulu noma, uma kwenzeka kuma-Mac, i-Thunderbolt 5 ene-RDMA.

Ngaphezu kwalokho, yize inani lamadivayisi amaningi aphansi lingafinyelela izibalo zenkumbulo ezihlaba umxhwele, Ngeke bahlale bencintisana ekusebenzeni okungahleliwe neqoqo lama-GPU aphezulu.Kukhona ibhalansi ebucayi phakathi kwezindleko zokuxhumanisa, ukubambezeleka kwenethiwekhi, kanye namandla okusebenzisa ikhompyutha asebenzayo okumele kucatshangelwe.

Ngayo yonke into, Isiphakamiso senani lika-Exo sisathakazelisa kakhulu: ukuguqula imishini ekhona, ngezindleko zehadiwe ezingenalutho, ibe yiqoqo elikwazi ukusebenzisa amamodeli abonakala sengathi akhethekile ezikhungweni zedatha ezikhethekile kuze kube muva nje.

Ukufakwa kwe-Exo, izidingo kanye nezindlela zokwenza

Ukuze i-Exo isebenze futhi isebenze ku-macOS Kuvamile ukusebenzisa amathuluzi afana ne-Homebrew ekuphathweni kwephakeji, kanye ne-UV ekuphathweni kokuxhomekeka kwe-Python, i-macmon yokuqapha ihadiwe kuma-chips e-Apple Silicon, i-Node.js yokwakha iphaneli yokulawula iwebhu, kanye ne-Rust (okwamanje esenguqulo yasebusuku) ekuhlanganiseni izibophezelo ezithile.

Umsebenzi ojwayelekile ku-macOS uyadlula shintsha indawo yokugcina esemthethweni ye-GitHubFaka ama-dependence bese usebenzisa umyalo ohambisanayo ukuze uqale i-Exo. Lokhu kuzoqalisa kokubili i-API kanye nedeshibhodi yewebhu efinyeleleka endaweni ekhelini http://localhost:52415Ukusuka lapho ungabona isimo seqembu, amamodeli, izimo, njll.

Ngaphezu kokwenziwa ngu Isikhumulo, I-Exo inikeza uhlelo lokusebenza lwendabuko lwe-macOS Lolu hlelo lokusebenza lusebenza ngemuva. Ludinga i-macOS Tahoe 26.2 noma kamuva futhi lusatshalaliswa njengefayela le-.dmg (isibonelo, ngaphansi kwegama elithi EXO-latest.dmg). Ngesikhathi sokufakwa, lungacela imvume yokushintsha izilungiselelo ezithile zesistimu nokwengeza iphrofayela yenethiwekhi, okudingekayo ukuze kusebenze izici ezithuthukisiwe njenge-RDMA.

Ezindaweni ze-Linux, ukufakwa kuvame ukuhilela ukusebenzisa umphathi wephakheji wesistimu. (njenge-apt ku-Debian/Ubuntu) noma nge-Homebrew ye-Linux uma uthanda. Khumbula ukuthi ezinye izinsiza, njenge-macmon, ziqondene ngqo ne-macOS futhi azisetshenziswa ku-Linux.

Okwamanje, Ku-Linux, i-Exo isebenza ku-CPUOkwamanje, akukho ukwesekwa okugcwele kwama-accelerator e-GPU. Ithimba lokuthuthukisa lisebenza ekwandiseni ukuhambisana namapulatifomu ahlukene ehadiwe, ngakho-ke kunconywa ukuthi uhlole izinkinga kanye nezicelo zezici ku-GitHub uma ufuna ukwazi ukuthi idivayisi yakho ethile ingasekelwa yini esikhathini esizayo.

Ukusetha i-RDMA phezu kwe-Thunderbolt ku-macOS

Esinye sezici ezintsha ezihlaba umxhwele kakhulu zezinguqulo zakamuva ze-macOS Kungukungezwa kokusekelwa kwe-RDMA, ikakhulukazi phezu kwamachweba e-Thunderbolt 5. Lesi sici sibalulekile ekunciphiseni ukubambezeleka phakathi kwama-Mac axhunywe ngokomzimba, okubaluleke kakhulu lapho kusatshalaliswa imodeli ye-LLM.

I-RDMA iyatholakala kusukela ku-macOS 26.2 kanye Isebenza kumadivayisi ane-Thunderbolt 5 njenge-M4 Pro Mac mini, i-M4 Max Mac Studio, i-M4 Max MacBook Pro, noma i-M3 Ultra Mac Studio. Kodwa-ke, esimweni esithile se-Mac Studio, kufanele ugweme ukusebenzisa i-Thunderbolt 5 port eseduze ne-Ethernet port, njengoba inemikhawulo ethile ye-RDMA.

Ukuze usebenzise i-RDMA, Kuyadingeka ukulandela inqubo kumodi yokutakulaIzinyathelo eziyisisekelo zihilela ukuvala i-Mac, ukuyivula ngokubamba inkinobho yamandla cishe imizuzwana eyi-10 kuze kuvele imenyu. ibhuthini, khetha inketho ethi “Izinketho” ukuze ufake indawo yokutakula bese uvula i-Terminal kusuka kumenyu ye-Utilities.

Ngaphakathi kwaleso siphetho sokutakula, Umyalo oboniswe kumadokhumenti e-Exo uyasetshenziswa ukuze unike amandla i-RDMA ezingeni lesistimu. Ngemva kokucindezela u-Enter nokuqala kabusha i-Mac, ukusebenza kuyavuselelwa futhi i-Exo ingayisebenzisa ukusungula izixhumanisi ze-latency eziphansi kakhulu phakathi kwama-Mac ahlukene axhunywe nge-Thunderbolt 5.

Ukusuka lapho, I-Exo iphatha ingxenye eyinkimbinkimbiUkusebenzisa i-RDMA ukunciphisa ukubambezeleka kokudluliselwa kwedatha phakathi kwama-node. Lokhu kuthuthukiswa kokuxhumana kwangaphakathi kubaluleke kakhulu lapho usebenza namamodeli aqukethe amakhulu ezigidigidi zamapharamitha ahlukaniswe emishinini eminingi.

Xhumana ne-Exo nge-API kanye nephaneli yewebhu

Ngaphezu kwedeshibhodi yezithombe ku-http://localhost:52415I-Exo iveza i-REST API ekuvumela ukuthi uphathe cishe yonke impilo yamamodeli: ukukhetha ukuthi uzowabeka kuphi, ukudala izimo, ukuqalisa izicelo zengxoxo, nokukhulula izinsiza lapho zingasadingeki.

Ukugeleza okuvamile kungaqala ngokubuza indawo yokugcina /instance/previewsUma uthumela amapharamitha emodeli ofuna ukuyisebenzisa, le nsizakalo iyabuya zonke izindawo ezivumelekile zalowo modeli eqenjini lakhoKungaba yinto efana nohlu "lwezinhlelo zokusabalalisa" ezingenzeka ngokusekelwe ku-topology yamanje kanye nezinsizakusebenza.

Ngokusekelwe empendulweni evela kulokho kubuka kuqala, Ukhetha isakhiwo esikufanele kakhulu. (isibonelo, eyokuqala ohlwini, ongayikhetha ngamathuluzi afana ne-jq kusuka ku-terminal) bese uyisebenzisela ukwenza isicelo se-POST endaweni yokugcina /isiboneloUmthwalo okhokhelwayo walesi sicelo kumele uvumelane nezinhlobo ezichazwe ku-CreateInstanceParams, kufaka phakathi amapharamitha afana nemodeli, ukucushwa kwehadiwe, kanye nama-node abelwe.

Uma isibonelo sesidaliwe, Ungaxhumana nayo usebenzisa i-endpoint ehambisana ne-OpenAI API, Ngokwesibonelo /v1/ingxoxo/ukuqedwaIsakhiwo sesicelo se-JSON sifana kakhulu naleso esisetshenziswa ne-ChatGPT API, okwenza kube lula kakhulu ukulungisa izinhlelo zokusebenza ezikhona ukuze zikhombe ku-Exo cluster yakho esikhundleni sefu.

Uma ungasadingi imodeli ethile, Ungayikhipha ngokuthumela isicelo se-DELETE nge-instance IDongakuthola ngokubuza ama-endpoints afana nalawa /isimo o /isiboneloNgakolunye uhlangothi, kuyasiza ukuba nezimpawu ezisheshayo, njengokuthi /amamodeli, ebala wonke amamodeli atholakala endaweni, noma umyalo we- /state uqobo, ochaza isimo samanje sokufakwa.

Isimo sephrojekthi kanye nomphakathi wokuthuthukiswa

I-Exo ibanjwe ku-GitHub ngaphansi kwe-repository exo-explore/exo Yakhiwe ngokusobala, ngeminikelo evela eqenjini eliyinhloko le-Exo Labs kanye nomphakathi. Ifayela le-CONTRIBUTING.md lichaza iziqondiso zokufaka isandla, ukuthumela izicelo zokudonsa, ukubika amaphutha, kanye nokuphakamisa izici ezintsha.

Kusukela Kungumsebenzi ophilayo, ohlala ushintshaKuvamile ngezinguqulo ezintsha, ukuthuthukiswa kokuhambisana namamodeli namapulatifomu ahlukene ehadiwe, kanye nezinguquko endleleni izingxenye ezithile ezilungiselelwe ngayo ukuze zikhishwe. Ngakho-ke, kuyalulekwa ukuthi ubuyekeze njalo amadokhumenti kanye nezinkinga ezivulekile.

Ngokuphathelene nokuhambisana nama-accelerator, i-macOS okwamanje ithola ukwesekwa okuthuthukisiwe ngenxa yokusebenzisa i-GPU ne-MLXNgesikhathi ikwi-Linux, okwamanje isebenza kwi-CPU. Ithimba likhuthaze abasebenzisi abanesifiso sokubona ukwesekwa kwezinhlobo ezintsha zehadiwe ukuvula noma ukusekela izicelo ezikhona ngezimpendulo ku-GitHub, ukuze bakwazi ukubeka phambili ukuthi yimaphi amapulatifomu adingeka kakhulu.

Ngokombono wezomnotho, I-Exo imahhala ngokuphelele futhi ivulekileAwekho amalayisense akhokhelwayo ahlobene nokusebenzisa isofthiwe, ehambisana kahle namaphrojekthi omuntu siqu, ezemfundo, noma amabhizinisi amancane afuna ukuzama ngempela i-AI ngaphandle kwesabelomali esikhulu.

I-Exo iziveza njengomdlali oyinhloko ekubuseni ngentando yeningi kwe-AI ethuthukisiwe.Lokhu kuvumela noma ubani onamadivayisi amaningi ekhaya noma ehhovisi ukuthi asungule iqoqo lakhe lamamodeli olimi. Nakuba kusenendlela okufanele ihambe, amathuba avulekayo maningi kakhulu okwenza izinto zisebenze ekhaya, abasizi bomuntu siqu, ukuhlaziywa kwedatha, kanye nezindawo zokuhlola zamamodeli anamuhla.