Impel ithuthukisa isipiliyoni se-Automotive Dealership isipiliyoni ngama-LLMS ahlelwe kahle ku-Amazon SageMaker

Lokhu okuthunyelwe kubhalwe phansi ngeTatia Tsmindashvili, Ana Kolkhidashvili, Guram Dentoshvili, Dachi Choladze kusuka Implel.

Impoqo iguqula izitolo zezimoto nge-AI-Powered Customent Customer Lifecycle Management Solution eshayela imisebenzi yokuthengisa kanye nokusebenzisana kwamakhasimende. Umkhiqizo wabo oyisisekelo, ukuthengisa i-AI, uhlinzeka ngokuzibandakanya kwamakhasimende okwenziwe ngohlobo lwezinsuku zonke, ukuphatha imibuzo eqondene nemoto kanye nemibuzo yezimoto yokuhweba kanye nezimali. Ngokufaka esikhundleni semodeli yazo yolimi lwesithathu okhona wesithathu (LLM) nge-Meta Llam Model ehlelwe kahle esetshenziswe ku-Amazon Sagemaker AI, impoqo ithole ukunemba okuthuthukile okungu-20% kanye nezilawuli zezindleko ezinkulu. Ukuqaliswa kusetshenziswa isici esibanzi esisethwe se-Amazon SageMaker, kufaka phakathi ukuqeqeshwa kwesisindo, i-AWQ-Aware yokuqonda isisindo (i-AWQ), kanye neziqukathi ezinkulu zemodeli (LMI). Le ndlela eqondene nesizinda ayithuthukisiwe kuphela ikhwalithi yemiphumela kodwa futhi ithuthukise ukuphepha nokusebenza okuqhubekayo okuqhathaniswa nama-LLMS ajwayelekile.

Kulokhu okuthunyelwe, sabelana ngendlela exakile ukuthi ithuthukisa kanjani isipiliyoni sezimoto sokuthengisa ngamakhasimende ngama-llms ahleliwe ku-sagemaker.

Ukuthengiswa kwe-Impl ai

I-Impl ikhulisa ukuthi abathengisi bezimoto baxhuma kanjani namakhasimende ngokuletha okuhlangenwe nakho okwenziwe ngezifiso kuzo zonke izifundo zokuthinta – kusuka ekuthengeni kokuqala ukuze uthenge, ngenkathi unikeza amakhono okwenziwa ngezinto ezenziwe ngezinto ezenziwe ngezimoto zokuxhumana kwamakhasimende. I-Sales Ai isebenzisa i-AI ekhiqizayo ukuhlinzeka ngezimpendulo ezisheshayo ubusuku nemini kumakhasimende azoba ngamakhasimende nangombhalo. Lokhu kuhlangana okugcinwe ngesikhathi sesigaba sokuqala sohambo lokuthenga lwekhasimende lokuthenga kwamakhasimende kuholela ekubunjisweni kwama-aphoyintimenti noma ukuxhumana okuqondile namaqembu okuthengisa. I-AI ye-AI inezici ezintathu eziyisisekelo zokuhlinzeka ngokubandakanyeka kwamakhasimende okungaguquki:

• Ukufingqa – Ukufingqa Ukuzibandakanya Kwamakhasimende Esedlule Ukuthola Inhloso Yekhasimende
• Isizukulwane sokulandela – Inikeza ukulandela okungaguquki kumakhasimende akhuthazelele ukusiza ukuvikela uhambo lokuthenga kwamakhasimende okufakiwe
• Ukuphendula Ukwenza okuthandwa nguwe – Izimpendulo zomuntu siqu zokuvumelanisa nemiyalezo yokuthengisa kanye nokucaciswa kwamakhasimende

Izici ezimbili ezibalulekile ziqhume ukuguqula uguquko kusuka kumhlinzeki wazo okhona we-LLM: isidingo sokwenza ngokwezifiso zemodeli kanye nokwenza kahle izindleko esikalini. Imodeli yamanani entengo yabo yangaphambilini yangemuva yaba yizindleko-njengoba ukuthengiselana amavolumu akhula, kanye nemikhawulo ekuhlelekeni okuhle kuvimbile ukuthi basebenzise ngokuphelele idatha yabo yokuphathelene ukuze bathuthukise amamodeli. Ngokusebenzisa imodeli ye-meta Lla ye-Meta Llama enhle ku-sagemaker, impoqo ithole okulandelayo:

• Ukuqagela izindleko ngamanani aphethwe, ukunciphisa izindleko ze-per-tokelen
• Ukulawulwa Okukhulu Ukuqeqeshwa Imodeli kanye nokwenza ngokwezifiso, okuholela ekuthuthukisweni kwama-20% ezicini eziyinhloko
• Ukucutshungulwa okuphephile kwemininingwane yokuphathelene ngaphakathi kwe-akhawunti yabo ye-AWS
• Ukulinganisa okuzenzakalelayo ukuze uhlangabezane ne-spike kwisidingo se-infence

Ukubuka konke

I-Impl ikhethe iSagemaker Ai, insizakalo yamafu ephethwe ngokuphelele eyakha, izitimela, kanye nezindlela zokufunda umshini (ML) amamodeli asebenzisa ingqalasizinda ye-AWS, amathuluzi, nokuhamba komsebenzi ukuze uhlele kahle i-Meta Lla Llam ye-AI AI. I-Meta Llama iyimodeli enamandla, elungele kahle imisebenzi eqondene nomkhakha ephathelene namandla aphezulu okulandela imiyalo, ukusekelwa kwamawindows anwetshiwe, nokusingathwa kahle kolwazi lwesizinda.

Impoqo esebenzisa i-sagemaker LMI iziqukathi zokufaka ukuboniswa kwe-LLM ngama-sagemaker endpoints. Lezi zitsheni zedokodo ezakhiwe ngenhloso zinikezela ukusebenza okwenziwe kahle kwamamodeli afana ne-meta llama ngokusekelwa kwamamodeli we-lora ahlelwe kanye ne-awq. I-Impel isetshenziswe i-lora enhle, inqubo esebenza kahle futhi engabizi kakhulu yokuhlehlisa i-LLMS yezinhlelo zokusebenza ezikhethekile, nge-Amazon SageMaker Studio Notebooks egijima kwi-ML.p4de.24xlarge Stances. Le ndawo ephethwe yenziwe lula inqubo yentuthuko, inika amandla i-Impel's Team ukuthi ihlanganise umthungo amathuluzi omthombo avulekile avulekile afana nePytorch ne-TorchPune yokuqeqeshwa ngemodeli. Okwe-Model Optimization, i-Implel isebenzise amasu we-AWQ ukunciphisa usayizi wemodeli nokuthuthukisa ukusebenza kwe-infence.

Emkhiqizweni, kushukumise ama-empoponds asetshenziswayo kuma-ML.G6E.12Xlarge times, anikwe amandla yi-NVIDIA GPUS enememori ephezulu, efanelekile ukusebenzela amamodeli amakhulu afana neMeta Llam ngempumelelo. I-Impel isebenzise isici se-sagemaker esakhelwe ngaphakathi ekulinganiseni okuzenzakalelayo ukukala ngokuzenzakalelayo iziqukathi ezisuselwa ezicelo ezifanayo, ezisiza ukufeza izimfuno zokukhiqiza eziguqukayo ngenkathi zisebenza ngezindleko.

Umdwebo olandelayo ukhombisa ukwakhiwa kwesixazululo, ukukhombisa imodeli ukuhleleka okuhle nokuthathwa kwamakhasimende.

Ithimba le-R & D le-Impl libambisene namaqembu ahlukahlukene ama-AWS, kufaka phakathi iqembu layo le-akhawunti, iqembu le-Genai Stustem, nethimba lesevisi yeSagemaker. Leli qembu elibonakalayo lasebenzisana ngaphezulu kwama-Sprints aholela ekutholeni okuhle kwe-Ai Lales AI Usuku lokubukeza ukuhlolwa kwemodeli, ukusebenza kwe-Benchmark SAGEMAKER PHELA, bese ulungiselela amasu wokulinganisa, bese ukhomba amasu we-sagemaker. Lokhu kubambisana okubandakanya amaseshini obuchwepheshe, imihlangano yokuqondanisa yamasu, kanye nezindleko nezingxoxo zokusebenza zokusebenza ngemuva. Ukubambisana okuqinile phakathi kwe-Indel kanye nama-AWS kwakutholile ekufezekiseni amandla aphelele wemodeli ephakeme ye-Finel.

Inqubo yokuhlola imodeli elungele kahle

Ukuguqulwa kwe-Impel ku-Meta Llama ye-Meta Llama elungiselelwe ukuthuthuka kuwo wonke amamethrikhi asemqoka ngokuthuthuka okubonakalayo ekuqondeni amagama akhethekile ezenzelwa izimoto futhi akhiqize izimpendulo ezenzelwe wena. Ukuhlolwa kwabantu okuhlelekile kuveze izithuthukisi ezindaweni ezibucayi zokuxhumana kwamakhasimende: izimpendulo ezenziwe ngezifiso zithuthuke kusuka kuma-73% kuye ku-86% ukunemba, ukugxuma kwesizukulwane semiyalezo ku-50%, kuqhamuke ku-59% kuya ku-92% ukunemba. I-screenshot elandelayo ikhombisa ukuthi amakhasimende axhumana kanjani ne-Sales AI. Inqubo yokuhlola imodeli efakwe empoqo yeqembu le-R & D Grading ahlukahlukene asetshenziswa amacala ahlinzekwa ngumhlinzeki we-LLM onamandla kanye namamodeli ahlelwe kahle.

Isibonelo sokusebenzisana kwamakhasimende nge-AI.

Ngaphezu kwekhwalithi yokukhipha, i-Implel elinganiselwe latency kanye nokusebenzisa ukuguqulela ukulungela ukukhiqizwa kwemodeli. Kusetshenziswa ama-AWScurl ngezicelo ze-HTTP ezisayiniwe ze-SIGV4, iqembu likuqinisekisile lokhu kuthuthukiswa kumametriki okusebenza kwangempela komhlaba, ukuqinisekisa okuhlangenwe nakho kwamakhasimende kahle ezindaweni zokukhiqiza.

Kusetshenziswa amamodeli aqondene nesizinda ukuze kusebenze kangcono

Ukuvela kwe-Impl Ukuvela kokuthengisa ai kuqhubekele phambili ku-LLM okujwayelekile kwe-LLM kuya kumodeli eqondene nesizinda, enhle. Kusetshenziswa idatha yokusebenzisana kwamakhasimende engaziwa, impoqo enhle ihlelwe imodeli yesisekelo esitholakala esidlangalaleni, okuholela ekuthuthukisweni okuningana okubalulekile. Imodeli entsha ikhombise ukukhuphuka kwama-20% ngokunemba kwezici eziyisisekelo, ukukhombisa ukuqondiswa kwezimoto ezithuthukisiwe zokuqonda kanye nokusetshenziswa kwewindows okusebenzayo. Ngokuguqukela kule ndlela, okuphoqelelayo kutholwe izinzuzo ezintathu eziyinhloko:

• Ukuvikeleka kwedatha okuthuthukisiwe ngokusebenza kwendlu ngaphakathi kwama-akhawunti abo ama-AWS
• Ukwehlisa ukuthembela kuma-API angaphandle kanye nabahlinzeki beqembu lesithathu
• Ukulawulwa Okukhulu Kokusebenza Kokulinganisa Nokwenza Ngokwezifiso

Lezi zithuthuka, ezihambisana nokwenza ngcono kwekhwalithi ebalulekile yekhwalithi, ukuguquguquka kwekhwalithi okuqinisekisiwe kuya kwimodeli ye-AI eqondene ne-AI eqondene ne-AI ye-AI ye-AI.

Ukunwebisa i-AI Innovation ekuthengisweni kwezimoto

Ukuphumelela kwe-Impl kusebenzisa amamodeli ahlelwe kahle ku-sagemaker kusungule isisekelo sokwengeza amakhono ayo e-AI ukusekela uhla olubanzi lwamacala asetshenzisiwe embonini yezimoto. I-Impel ihlela ukuguqukela kumamodeli asendlini, eqondene nesizinda ukwelula izinzuzo zokunemba okuthuthukile kanye nokuhamba phambili komkhiqizo wazo we-AI ngokutholwa kokubuyiselwa kwemali okutholakalayo, ukushayela umsebenzi okuthuthukile, ukushayela umsebenzi we-agentic. Lokhu okusha kungasiza ukuletha izinhlelo eziguqukayo, zokwazi isimo eziklanyelwe ukuxhumanisa, ukucabanga, kanye nokwenza imisebenzi yazo zonke izinto zezezimoto ezibuyisanayo.

Ukugcina

Kulokhu okuthunyelwe, sixoxile ukuthi i-Endel isithuthukise kanjani isipiliyoni samakhasimende ezithunyelwayo nge-LLMS elungisiwe ku-sagemaker.

Izinhlangano ezibheka izinguquko ezifanayo kumamodeli ahlelwe kahle, okuhlangenwe nakho okuyisisekelo kukhombisa ukuthi ukusebenza nge-AWS kungasiza kanjani ukufeza ukuthuthuka kokunemba kanye nama-Model ai calls ngenkathi enza amakhono wesikhathi eside we-AI ahambelana nezidingo ezithile zomkhakha. Xhuma nethimba lakho le-akhawunti noma vakashela i-Amazon Sagemaker AI ukuze ufunde ukuthi i-sagemaker ingakusiza kanjani ukufaka futhi uphathe amamodeli ahlelwe kahle.

Mayelana nababhali

Nicholas scozzafiva Ingabe ukwakhiwa kwezixazululo eziphezulu ku-AWS, kugxile kumakhasimende wokuqalisa. Ngaphambi kokuba nendima yakhe yamanje, wasiza amakhasimende amabhizinisi azulazule uhambo lwawo lwefu. Unothando ngengqalasizinda yefu, ezenzakalelayo, ama-devops, nokusiza amakhasimende ukwakha nokukala kuma-AWS.

Sam sudakoff ungumphathi we-akhawunti ophezulu ngama-AWS, agxile kumasu wokuqalisa amaqhinga. USam ubheka phambili emathafeni e-Technology, AI / ML, kanye nezisombululo ze-AWS. Uthando lukaSam lulele ekuqaleni kokuqala nokushayela i-SAS kanye ne-AI Transformtions. Ngokuphawulekile, umsebenzi wakhe onama-aws aphezulu wokuqalisa ama-ISV agxile ekwakheni ubudlelwano obuhle kakhulu ekwakheni ubuchwepheshe bebhizinisi le-Bridge Enterprise ngezixazululo zokuqala ezintsha, ngenkathi ugcina ukubambelela ngokuqinile ngokuphepha kwedatha nezidingo zobumfihlo.

Vivek gangasani Ingabe ukwakhiwa kwezixazululo ezikhethekile zokutholwa kokutholwa kuma-AWS. Usiza izinkampani ezivelayo ze-AI ezivelayo zakha izixazululo ezintsha zisebenzisa izinsizakalo ze-AWS kanye ne-ashevuzelo compute. Njengamanje, ugxile ekuthuthukiseni amasu okuhleleka okuhle nokwenza kahle ukusebenza kokuqashwa kwamamodeli amakhulu olimi. Esikhathini sakhe samahhala, i-Vivek ikujabulela ukuhamba ngezinyawo, ukubuka ama-movie, nokuzama ama-cuisines ahlukile.

UDmitry waklama Ingabe ukwakhiwa kwe-AI / ML solutions okuphezulu kwe-AI / ML kuma-AWS, kusiza amakhasimende aklama futhi akhe izixazululo ze-AI / ML. Umsebenzi we-Dmitry uhlanganisa uhla lwamacala okusetshenziswa kwe-ML, ngentshisekelo eyinhloko ekwenzeni i-AI eqinile, ukufunda okujulile, kanye nokulinganisa ml ngaphesheya kwebhizinisi. Uye wasiza ezinkampanini ezimbonini eziningi, kufaka phakathi umshuwalense, izinsizakalo zezezimali, izinsiza, kanye nokuxhumana ngocingo. Ngaphambi kokujoyina ama-AWS, uDmitry wayengumdwebi wezakhiwo, unjiniyela, kanye nobuholi bezobuchwepheshe kuma-Data Analytics kanye nezinkundla zokufunda ngomshini embonini yezimali zezezimali.

Tatia tsmindashvili ungumcwaningi omkhulu wokufunda ojulile ku-Implel ne-MSC eBioomedical Engineering kanye nemininingwane yezokwelapha. Uneminyaka engaphezu kwengu-5 yesipiliyoni e-AI, enezintshisakalo ezihamba ama-LLM agents, ukumbumbuluzwa, kanye ne-neuroscience. Ungamthola ku-LinkedIn.

Ana kolkhidashvili Ingabe umqondisi we-R & D empondweni, lapho ahola khona imizamo ye-AI egxile kumamodeli amakhulu olimi kanye nezinhlelo zengxoxo ezenzakalelayo. Uneminyaka engaphezu kwengu-8 yesipiliyoni e-AI, enakekela amamodeli amakhulu ezilimi, izinhlelo ezizenzakalelayo zengxoxo, ne-NLP. Ungamthola ku-LinkedIn.

Guram dentoshvili Ingabe umqondisi we-Engineering no-R & D kwimpoqo, lapho ahola khona ukuthuthukiswa kwezixazululo ze-AI ezihlelekile futhi zishayela izinto ezintsha emikhiqizweni yenkampani ye-AI. Waqala umsebenzi wakhe ePulsar Ai njengonjiniyela wokufunda umshini futhi wadlala indima ebalulekile ekwakheni ubuchwepheshe be-AI ehambelana nemboni yezimoto. Ungamthola ku-LinkedIn.

Dachi choladze Ingabe iSikhulu Esiyisikhulu Esithuthukile kwi-Impl, lapho aholela khona imizamo ngecebo le-AI, emisha, kanye nentuthuko yomkhiqizo. Uneminyaka engaphezu kwengu-10 yesipiliyoni esonweni lamabhizinisi nobuhlakani bokufakelwa. UDachi ngumsunguli we-Pulsar Ai, ekuqaleni kwe-AI Startup yaseGeorgia, okwakuhlanganiswa kamuva nge-Impl. Ungamthola ku-LinkedIn.

UDeepAm Mishra Ingabe umeluleki we-SR wokuqalisa ama-AWS futhi weluleka ukuqala kwe-ML, ukuphepha kwe-AI, kanye nokuphepha kwe-AI. Ngaphambi kokujoyina ama-AWS, i-PEEPAM yasungulwa futhi yahola ibhizinisi le-AI kwaMicrosoft Corporation kanye ne-Wipro Technologies. UPeepam ubelokhu esomabhizinisi osolwa ngokweserial kanye nomtshali zimali, esungule ama-4 AI / ML Startups. I-Deepam isuselwa endaweni ye-NYC Metro futhi ijabulele ukuhlangana nabasunguli be-AI.