Kwenzekeni
I-Deepgram imemezele izengezo ezimbili ezibonwayo zezinhlobo zayo zokukhuluma-kuya-umbhalo kanye nombhalo-ube-inkulumo ezisetshenziswe njenge-Amazon SageMaker AI yesikhathi sangempela sokuphela. Ama-Metrics athuthukisiwe ashicilela idatha yokusetshenziswa neyokukhokha ku-akhawunti yekhasimende ye-CloudWatch, kuyilapho ukuhlanganiswa kwe-Prometheus ne-OpenTelemetry kuveza injini, umsingathi kanye nezilinganiso ze-GPU ngayinye nge-SageMaker AI yokubonakala enemininingwane.
Isimemezelo sika-Deepgram siphathelene namamodeli wenkulumo-umbhalo kanye nombhalo-ube-inkulumo asebenza njengamaphakheji emodeli ye-Marketplace Amazon SageMaker AI yesikhathi sangempela ku-akhawunti yekhasimende AWS. Inkampani ithi okulalelwayo nokulotshiwe kuhlala kuleyo akhawunti kuyilapho i-SageMaker ihlinzeka ngezilawuli zokuthunyelwa, zokukala nokuqapha. Okuthunyelwe kufaka amakhono amasha njengokuvala igebe lokubonakala: ukuqapha okujwayelekile kwendawo yokugcina kungabonisa ukutholakala nokuphathwa kwesicelo, kodwa kungase kungavezi ukuthi yiziphi izici zenkulumo ezisetshenziswayo, yimaphi amayunithi ashayela amashaji emakethe noma ukuthi inkomba isebenzisa kanjani i-GPU ngayinye.
Isengezo sokuqala, esibizwa ngokuthi i-Deepgram Enhanced Metrics, sithumela ulwazi lokukhokha nelokusetshenziswa ku-Amazon CloudWatch ngamarekhodi e-CloudWatch Embedded Metric Format abhalwe ekuphumeni okujwayelekile kwesiqukathi. I-SageMaker idlulisela lokho okukhiphayo eqenjini lokungena le-CloudWatch le-endpoint, lapho i-CloudWatch ikhipha amarekhodi abe amamethrikhi ajwayelekile. Ngokomthombo, le nqubo ayidingi i-ejenti ehlukile, inqola eseceleni noma imvume eyengeziwe ye-IAM futhi isebenza ngokuhlukaniswa kwenethiwekhi ye-AWS Marketplace ngoba isebenzisa indlela yokungena ekhona ye-SageMaker-to-CloudWatch. Indawo yamagama ekhokhayo ithi Deepgram/SageMakerInference. Imethrikhi yayo ye-ConsumedUnits imelela amayunithi okukhomba akhokhisekayo okusakaza-bukhoma okuqediwe, okurekhodiwe kusengaphambili kanye nezicelo zokuguqula umbhalo ube inkulumo, kuyilapho i-AudioDurationSeconds ne-CharCount zichaza umsindo ocutshunguliwe nezinhlamvu ezihlanganisiwe. I-Deepgram ithi lawa manani afanayo nalawo asetshenziswa ekukhokhisweni kwemitha ye-AWS Marketplace.
Ukusakazwa kokusetshenziswa kwesibili, i-Deepgram/SelfHosted, ikhishwa iseva ye-Deepgram API futhi ibonisa indlela ithrafikhi eyisebenzisa ngayo isevisi kunokubuyisana ngokuqondile ibhili. Umthombo ubala amamethrikhi okusakaza-bukhoma nevolumu erekhodiwe, amamodeli amamodeli afana ne-nova-3 kanye ne-flux, nezici ezihlanganisa ukufakwa kwe-dayari, ukufometha okuhlakaniphile, ukuphindaphinda kanye nokwaziswa kwethemu elingukhiye. Lawa mamethrikhi ahlanganisa kuzo zonke izindawo zokugcina ze-Deepgram ku-akhawunti ye-AWS kanye Nesifunda. Okuthunyelwe kuthi basebenzisa ubukhulu bekhadinali ephansi ngaphandle kwemibhalo ebhaliwe, okokufaka kombhalo ube inkulumo noma izihlonzi zesicelo ngasinye, kodwa azifaki igama lephoyinti lokugcina noma i-ID yesibonelo. Ukusakaza kokusetshenziswa kungakhutshazwa ngokukhipha okuguquguqukayo kwemvelo, kuyilapho ukusakaza kokukhokha akukwazi ukukhutshazwa ngoba i-Deepgram ikuchaza njengengxenye yepayipi lokulinganisa.
Ngokusebenza kwezinga eliphansi, i-Deepgram ithi iziqukathi zayo zidalula amamethrikhi e-Prometheus okungaqoqwa ukubonwa okunemininingwane kwe-SageMaker AI nge-OpenTelemetry Collector ephethwe yi-AWS ephethwe yi-OpenTelemetry egijima endaweni ngayinye yokugcina. Idatha ewumphumela ihlanganisa amamethrikhi okuthunyelwa kwamanye amazwe e-DCGM ye-GPU ngayinye, ama-metrics we-node-exporter we-CPU engumsingathi nenkumbulo, kanye namamethrikhi enjini ye-Deepgram njengezicelo zokusakaza ezisebenzayo kanye nomthamo wokusakaza olinganiselwe. Umthombo uthi uchungechunge ngalunye luphethe amalebula esisetshenziswa se-SageMaker, afaka iphoyinti lokugcina, okuhlukile nezihlonzi zezibonelo. Amakhasimende angabuza indawo ethile yokugcina, isibonelo noma i-GPU esebenzisa i-PromQL nge-CloudWatch, Grafana noma elinye ithuluzi elihambisanayo. Ukubonakala okunemininingwane kunikwe amandla ngokuzenzakalela kumaphoyinti okugcina asanda kwakhiwa, anemvamisa yokushicilela engamasekhondi angu-60, kuyilapho iziphetho ezindala zidinga isibuyekezo sokucushwa kwephoyinti lokugcina kusetshenziswa ukusetshenziswa okuluhlaza okwesibhakabhaka/okuluhlaza.
Imininingwane yomthombo: aws.amazon.com ↗
Kungani kubalulekile
Izinguquko ziqondise inkinga engokoqobo ku-AI esingethwe yona: amakhasimende angakwazi ukuqapha ukuthi indawo yokugcina iyasebenza yini, kodwa angase angabonakali kumodeli yokusetshenziswa kokushayela, amayunithi angemuva kokukhokhiswa kwemakethe kanye nomthamo wamasheshisi angawodwana. Umthombo uchaza indlela yokuxhuma leyo mibuzo yokusebenza nezezimali ngaphandle kokuvula ukufinyelela kwenethiwekhi okuphumayo kusukela kusiqukathi esiyimodeli.
Ukubaluleka okusheshayo ukubonakala kokusebenza. Ukuqondiswa kwenkulumo akuwona umthwalo wokusebenza ofanayo: izikhathi zokusakaza, umsindo orekhodiwe kusengaphambili kanye nezicelo zokuguqula umbhalo ube-inkulumo zingabeka izimfuno ezihlukene ekusetshenzisweni, futhi izici ozikhethela zingashintsha ivolumu yokucubungula noma ukusetshenziswa kwensiza. Amamethrikhi okusetshenziswa kweleveli ye-akhawunti ye-Deepgram ahloselwe ukubonisa lowo mehluko ngokwemibandela opharetha namathimba ezezimali angasebenzisa. Ikhasimende lingahlola ukuthi ingakanani ithrafikhi esebenzise ukudayela noma liqhathanise ukusetshenziswa kwe-model-tier ngokuhamba kwesikhathi. Lokho kungasiza izinhlangano zihlonze amaphethini okusetshenziswa angalindelekile, zenze amasistimu okubuyisela emuva angaphakathi noma zinqume ukuthi imiphi imithwalo yomsebenzi okufanele ihanjiswe ekusetshenzisweni okuhlukile. Umthombo wethula lokhu njengokusetshenziswa kwamamethrikhi, hhayi ukonga okubonisiwe noma ukuthuthukiswa kokusebenza.
Uxhumano lwenkokhelo lungenza ukuthengwa kwe-AI okusekelwe emakethe kube lula ukucwaningwa. I-Deepgram ithi i-metric ye-ConsumedUnits iphethe amanani eyunithi ekhokhisekayo afanayo asetshenziswa kumitha ye-AWS Marketplace, okuvumela amakhasimende ukuthi aqhathanise isamba se-CloudWatch nenkokhiso yawo ye-AWS. Lokho kucace kakhulu kunesibalo sesicelo, ngoba isicelo singase simele iseshini yokusakaza-bukhoma, inani lomsindo noma ivolumu yezinhlamvu ezihlanganisiwe. Umthombo uthi izici ezijwayelekile ze-CloudWatch—okuhlanganisa amadeshibhodi, ama-alamu kanye nezibalo zemethrikhi—zingasetshenziswa kudatha. Ayinikezi isibonelo somphumela wokubuyisana, isilinganiso samaphutha, isilawuli sezimali noma isiqinisekiso esizimele sokuthi amamethrikhi ahlala efana ne-invoyisi. Izinhlangano zisazodinga amasheke azo ezezimali kanye nokuthobela imithetho.
I-GPU ngayinye kanye ne-metrics yezinga lenjini ibhekana nenkinga ehlukile: ukuhlela umthamo ekusetshenzisweni okufakwe esikalini. Isilinganiso esibanzi semikhumbi singasitha isisheshisi esilayishwe kakhulu, kuyilapho ukubalwa kwesicelo kukodwa kungase kungabonisi i-headroom yokusakaza ngesikhathi esisodwa. I-Deepgram ithi umthamo wokusakaza wenjini yayo ungaqhathaniswa nezicelo ezisebenzayo zokuhlinzeka ngesignali ebikwe injini ngezinqumo zokukala, nokuthi amamethrikhi e-GPU angahlolwa ngokwehlukana ezimweni eziningi ze-GPU. Lokhu kungasiza opharetha baphenye ukusetshenziswa okungalingani, bashune ama-threshold e-autoscaling noma banqume ukuthi uhlobo lwesibonelo luba yini umgoqo. Amagama abalulekile: inani lamandla liwukulinganisa kwenjini, akusona isiqinisekiso esiqinisekisiwe esizimele sokuthi mangaki ama-streams isenzakalo esiyoqhubeka ngaphansi kwayo yonke ifomethi yomsindo, imodeli, inhlanganisela yesici noma iphethini yomthwalo womsebenzi.
I-engeli yezokuphepha nokubusa nayo iqinile. Umthombo uthi iziqukathi ze-Makethe ye-AWS zisebenza ngokuhlukaniswa kwenethiwekhi futhi azikwazi ukwenza uxhumano oluphumayo, idizayini ekhethwe amakhasimende athile ngenxa yezizathu zokuphepha nokuhambisana. Indlela ye-telemetry ehlongozwayo ye-Deepgram igcina izilinganiso ngaphakathi kwendawo yekhasimende ye-CloudWatch ngaphandle kokudinga ukuthi isiqukathi sivule umzila wenethiwekhi. Umthombo uphinde uthi izilinganiso ezisohlwini aziqukethe ulwazi lomuntu siqu olungamhlonza futhi akubandakanyi imibhalo noma izihlonzi zesicelo. Lezo zitatimende zichaza izakhiwo ezimenyezelwe, hhayi ukuhlola okuphelele kobumfihlo: amakhasimende ahlala enesibopho sokucushwa kwawo kwe-AWS, izilungiselelo zokugcinwa, izilawuli zokufinyelela kanye nezibopho zokulawula ngaphansi kwemodeli yesibopho esabiwe.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Umthombo awunikezi izilinganiso ezizimele zokuqapha ngaphezulu, ukonga izindleko, ukunemba kwezilinganiso zamandla noma ukwamukelwa kwamakhasimende. Amaqembu acabangela ukusetshenziswa azodinga ukuhlola ukuthi ingabe amamethrikhi okukhokhiswa kweleveli ye-akhawunti anembudumbudu ngokwanele yini, ukuthi izilinganiso zenjini ziyafana yini nethrafikhi eboniwe nokuthi yiziphi izindleko ezengeziwe ze-CloudWatch, ukugawula nokusingathwa kwe-GPU eziphumela ezindaweni zazo.
Okokuqala okungaziwa ukusebenza komhlaba wangempela. Isimemezelo asibiki i-CPU, inkumbulo, isitoreji noma ukubambezeleka okungaphezulu kokubhala kwamamethrikhi ashumekiwe, ukuklwebheka amaphoyinti e-Prometheus noma ukushicilela idatha njalo ngamasekhondi angu-60. Futhi ayiwabali amanani ezindleko ze-CloudWatch zamalogi namamethrikhi, izindleko zabaqoqi nanoma yimuphi umthelela ekuziphatheni kokuskena okuzenzakalelayo. I-AWS kanye ne-Deepgram inothi lokuthi ukusingathwa kwe-endpoint, izimo ze-GPU, amalogi namamethrikhi e-CloudWatch, ingqalasizinda yenethiwekhi nengqalasizinda ehlobene konke kungaba nezindleko, ngisho naphakathi kwesivivinyo samamodeli esishiwo sezinsuku eziyi-14 se-Deepgram. Abasebenzisi abangase babe ngabasebenzisi bazodinga izilinganiso eziqondene nomthwalo womsebenzi kunokuthi bacabange ukuthi ukubonwa okuthuthukisiwe akuhlanganisi izindleko.
Inkinga yesibili i-metric granularity. Ukukhokhisa nokusetshenziswa kokusetshenziswa kwe-Deepgram kuhlanganisa ezindaweni zokuphela kwe-akhawunti neSifunda futhi akubonisi isiphetho noma isenzakalo. Lokho kungase kwanele ekubikeni okubanzi kwezezimali, kodwa ngeke ngokwako kuphendule ukuthi yikuphi ukusetshenziswa okukhiqize umthamo othile. Ukuhlaziywa kwe-per-end kanye ne-GPU ngayinye kuncike endleleni ehlukene ye-Prometheus kanye ne-OpenTelemetry kanye nasekubonweni okuningiliziwe okunikwa amandla. Umthombo uthi amaphoyinti okugcina amasha anokuthi izilungiselelo ezinikwe amandla ngokuzenzakalelayo nezindala zingabuyekezwa, kodwa azichazi amacala onqenqema lokufuduka, izikhathi zokugcinwa, izifanekiso zedeshibhodi noma ukuthi opharetha kufanele baluphathe kanjani uchungechunge olulahlekile, olubambezelekile noma olungqubuzanayo.
Isignali yamandla nayo ifanelwe ukuqinisekiswa. I-Deepgram ichaza i-engine_estimated_stream_capacity njengokulinganisa kwenjini kokusakazwa okufanayo okuzinzile. Umthombo awubonisi ukuthi leso silinganiso sibalwa kanjani, ukuthi sishintsha kanjani ngesigaba semodeli noma izici ezinikwe amandla, noma ukuthi sisebenza kanjani phakathi nokukhuphuka kwethrafikhi nezimo ezonakalisiwe. Abasebenzisi kufanele bayiqhathanise nokubambezeleka okuphawuliwe, amaphutha, umugqa kanye namaseshini aqediwe ngaphambi kokuyisebenzisa njengesicuphi sokukala esizenzakalelayo. Amamethrikhi ajwayelekile e-SageMaker afana ne-ConcurrentRequestsPerModel, FirstChunkLatency kanye ne-Invocation5XXErrors ahlala ehambisana ngoba ukusakaza okusha kuyahambisana kunokuthatha indawo yakho.
Okokugcina, isimemezelo sisungula ukutholakala kokuthunyelwa kwe-Deepgram SageMaker AI, kodwa hhayi umthelela obanzi we-ecosystem. Ayisho ukuthi mangaki amakhasimende asebenzise amamethrikhi, noma ngabe abanye abathengisi benkulumo-AI banikeza ukubonakala okufanayo, noma ukuthi i-AWS izonweba yini ukuhlanganiswa okufanayo kumaphakheji emodeli engeziwe. Ukulandelela okungokoqobo ukuthi abasebenzisi bangakwazi yini ukuvumelanisa ama-invoyisi ngendlela enokwethenjelwa, bahlukanise izindawo eziyizisetshenziswa futhi balawule ukusetshenziswa kwezinga lesici ngaphandle kokwengeza ingqalasizinda entsha. Ubufakazi obuvela ekusetshenzisweni okuzimele, ukuziphatha kwemethrikhi okurekhodiwe ezikhathini ezinde nolwazi lwamakhasimende buzocacisa ukuthi isici sishintsha ngokuphathekayo ukubusa kwansuku zonke kwenkulumo esingethelwe yona i-AI.