Hacker News

Gukomeza guterana uhereye kumahame ya mbere (2025)

Gukomeza guterana uhereye kumahame ya mbere (2025) Isesengura ryuzuye ryikomeza ritanga isuzuma rirambuye ryibice byingenzi hamwe ningaruka nini. Ibice by'ingenzi byibandwaho Ibiganiro byibanda kuri: Uburyo bwibanze na ...

6 min read Via huggingface.co

Mewayz Team

Editorial Team

Hacker News

Gukomeza Gufata Amahame Yambere (2025)

> Kubisobanukirwa uhereye kumahame ya mbere byerekana impamvu yahindutse imyubakire yifatizo kuri buri sisitemu yo mu rwego rwo hejuru ikora ya AI ikoreshwa ku gipimo cya 2025.

Niki Mubyukuri Gukomeza Gukomeza kandi Kuki Gufata Igihagararo Byatsinzwe?

Kugirango ushimishe guhora, ugomba kubanza kumva icyo wasimbuye. Amatsinda gakondo yo gutondekanya amatsinda ahamye yumubare wibyifuzo hamwe, abitunganya nkigice kimwe, kandi yemera gusa ibyifuzo bishya nyuma yicyiciro cyose kirangiye. Inenge ikomeye ni uko imvugo nini yerekana ururimi rutanga ibimenyetso byuburebure - icyifuzo kimwe gishobora kurangira nyuma yikimenyetso 20 mugihe ikindi mugice kimwe gikora 2000. Buri GPU muri cluster yicaye ubusa itegereje urukurikirane rurerure kugirango irangire mbere yuko umurimo mushya utangira.

> Ikora kurwego rwa itera aho gukora urwego rusaba. Nyuma ya buri kimwe cyimbere cyanyuze mucyitegererezo, gahunda itegura niba urutonde urwo arirwo rwose rugeze ku ndunduro yacyo. Niba ifite, icyo kibanza gihita gisubirwamo kandi gihabwa icyifuzo cyo gutonda umurongo - nta gutegereza, nta myanda. Ibice bigize ibice bihindagurika neza na buri ntambwe ya decode, kugumya gukoresha ibikoresho hafi ya theoretical ntarengwa igihe cyose.

Nigute Cache Cache ikorana nogukomeza guhora kurwego rwa sisitemu?

Urufunguzo-agaciro cache nuburyo bwo kwibuka butuma impinduka zifatika zifatika. Kuri buri kimenyetso cyatunganijwe, icyitegererezo kibara urufunguzo rwo kwitondera nindangagaciro zigomba kugumana bityo ibimenyetso byakurikiyeho ntibisubiremo kubara. Muri sisitemu yo gutondeka neza, kugabura KV kugabanwa biroroshye: kubika ububiko bujyanye nuburebure ntarengwa bwakurikiranye kuri buri cyifuzo mu cyiciro.

Gukomeza guterana bigora ibi neza. Kuberako ibyifuzo byinjira kandi bisohoka mugice mugihe kitateganijwe, sisitemu ntishobora kubanza kugabura ibintu byahagaritswe kwibuka. Iyi niyo mpamvu rwose vLLM's PagedAttention - yatangijwe mumwaka wa 2023 - yabaye ntaho itandukaniye no gukomeza guhora mubikorwa byoherejwe. Urupapuro rwitondewe ruguza moderi yububiko bwa paji yububiko bwa sisitemu ikora, igabanya Kache cache mubice bidahuza ubunini bungana. Urupapuro rwihishwa rwurupapuro rushobora gukwirakwira mububiko bwa GPU nkuko paji yibuka yibintu ikwirakwizwa muri RAM igaragara. Igisubizo kiri hafi ya zeru yibuka imyanda ivuye mu gucamo ibice, igahita ihinduranya ingano nini yo mu rwego rwo hejuru hamwe n’ibisohoka byinshi nta shoramari ry’inyongera.

Nubuhe buryo Bwingenzi bwo Guteganya Gukora Gukomeza Gukora?

Ibyemezo bitatu byuzuzanya byateganijwe bigenga buri sisitemu ikomeza:

    • Swap ishingiye kubanza kubika ibarwa ariko ikoresha umurongo wa PCIe; kubara bitesha agaciro GPU ariko bikomeza kwibuka neza.
    • Igenzura ryinjira: Gahunda igomba guteganya niba icyifuzo gishya cya KV cache kizahuza nibiboneka mugihe cyacyo cyose. Gusuzugura bitera impanuka-yibuka impanuka hagati; gukabya kwicisha inzara umurongo bitari ngombwa. Sisitemu zigezweho zikoresha uburebure buringaniye bwo kugabura no kubika ibicuruzwa kugirango uhuze izo ngaruka.
    • Ibice byacitsemo ibice bigabanije ibisobanuro birebire mubice bingana-bifatanye hamwe na decode itera, bigabanya igihe-kuri-mbere-ikimenyetso-cyihuta cyihuta kubakoresha bahurira ku giciro cyo hasi y’ibicuruzwa byinjira mbere.
    • Gutonda umurongo: Ibikorwa byoherejwe na sisitemu icyiciro cya SLA. Ubukererwe-bwita API bwita kubanza-imbaraga zumurimo wakazi. Hatariho iki gice, inyandiko imwe ndende yo gukusanya incamake irashobora gutesha agaciro ubunararibonye bwabakoresha kumagana amagana ahuriweho.
    ""

    Nigute Kohereza-Isi-Isi Yapima Inyungu Yakozwe?

    Ibipimo ngenderwaho biva muri Anyscale, hamwe n’imyororokere yigenga mu miryango myinshi y’icyitegererezo mu 2024, ihora yerekana ibyiciro bikomeza bitanga ibicuruzwa biva hagati ya 23 × na 36 × byinjira cyane ugereranije na nave static batching ukurikije inzira nyabagendwa. Inyungu zigaragara cyane mugihe ibyifuzo birebire bitandukanijwe ni byinshi - mubyukuri ibintu biranga umusaruro uganira kubikorwa bya AI akazi aho ibibazo byabakoresha bitangirira kumagambo atatu yatanzwe kugeza kumpapuro nyinshi zoherejwe.

    💡 DID YOU KNOW?

    Mewayz replaces 8+ business tools in one platform

    CRM · Invoicing · HR · Projects · Booking · eCommerce · POS · Analytics. Free forever plan available.

    Start Free →

    Ubukererwe buvuga inkuru irenze. Igihe-kuri-mbere-token iratera imbere kuburyo bugaragara kuko sisitemu itagitegereje ko icyiciro cyuzuye giteranya mbere yo gutangira prefill. Ubukererwe hagati ya token buguma butajegajega munsi yumutwaro uringaniye ariko butesha agaciro muburyo bwuzuye aho kugwa, kuko gahunda ikomeza gutera imbere kumurongo wose ukora nubwo umurongo ukura cyane. Kubucuruzi bwubaka igihe nyacyo kiranga AI, iyi ntera nziza yo gutesha agaciro akenshi irahambaye mubucuruzi kuruta umubare winjiza cyane.

    Nigute Abashoramari Bashobora Gukurikiza Amahame Yogukomeza Kurenga AI?

    > Sisitemu y'imikorere yubucuruzi ihura ningorabahizi imwe: imirimo yigihe cyinyamanswa ihatanira ubushobozi bwo gutunganya ibintu murwego rwa CRM, gutangiza ibicuruzwa, imiyoboro yisesengura, hamwe nibikorwa bya e-bucuruzi.

    Mewayz akoresha iyi filozofiya muri OS yayo 207-module yubucuruzi, ikoresha imbaraga zimirimo ikora kumurongo uhuriweho ukoreshwa nubucuruzi 138.000 kwisi yose. Aho guhatira amakipe gutegereza icyiciro cyo gutanga raporo, umurongo wemejwe, cyangwa ibikoresho byacecetse, Mewayz atunganya ibikorwa byubucuruzi ubudahwema - kugaburira ibisubizo byuzuye ahita yinjira mumasoko yuburyo buryo gahunda yo gukomeza guterana ibyokurya bya GPU bisubizwa kumurongo wabisabye. Igisubizo kirapimwa kunonosora ibicuruzwa mubikorwa byubucuruzi, ntabwo ari ibipimo gusa.

    Ibibazo bikunze kubazwa

    Gukomeza guterana kimwe no gukinisha imbaraga muri TensorFlow Gukorera?

    Oya. TensorFlow Serving ifite imbaraga zo guteranya ikusanya ibyifuzo mubice byubunini buhindagurika ukurikije igihe windows hamwe nuburebure bwumurongo, ariko iracyatunganya buri cyiciro atomike kuva itangiye kugeza irangiye. Gukomeza gutondeka bikorera kumurongo wikimenyetso cyintambwe, kwemerera icyiciro cyo guhindura inzira yose yimbere. Itandukaniro rya granularité niyo mpamvu gukomeza guterana bigera kumurongo mwinshi cyane kubikorwa byimikorere ya autoregressive yimirimo yihariye.

    Gukomeza guterana bisaba guhindura imiterere yubwubatsi?

    Ububiko busanzwe bwa transformateur ntibisaba guhinduka. Gukomeza gutondeka bishyirwa mubikorwa murwego rwo gutanga serivisi binyuze mumahinduka kuri gahunda yerekana, umuyobozi wibuke, hamwe na kernel. Nyamara, bimwe muburyo bwiza - cyane cyane Urupapuro rwitonderwa - bisaba intungamubiri za CUDA zisimbuza gushyira mubikorwa bisanzwe, niyo mpamvu umusaruro-urwego rwo gukomeza guteranya ibintu nka vLLM na TensorRT-LLM ntabwo ari ugusimburwa kubisimbuza rusange-intego zifatika.

    Ni izihe mbogamizi zibyuma bigabanya gukora neza?

    Umuyoboro mugari wa GPU HBM hamwe nubushobozi bwa VRAM nimbogamizi zibanze. Kinini nini ya KV isaba kwibuka cyane, igabanya guhuza byinshi. Umuyoboro mwinshi cyane (NVLink, Infiniband) uba ingirakamaro kubikorwa byinshi bya GPU aho cache igomba gukwirakwizwa mubikoresho. Mubidukikije bibujijwe kwibuka, kugereranya ubukana bwa KV cache indangagaciro (kuva FP16 kugeza INT8 cyangwa INT4) igarura ubushobozi ku giciro cyo gutesha agaciro kworoheje kwemerwa mubikorwa byinshi byubucuruzi.


    Waba wubaka ibiranga imbaraga za AI cyangwa utegura ibikorwa byubucuruzi bigoye mumuryango wawe wose, ihame shingiro rirasa: gukuraho igihe cyubusa, kugarura ubushobozi ubudahwema, no gutunganya imirimo myinshi hamwe nubutunzi ufite. Mewayz ashyira mu bikorwa iryo hame mu bice 207 byahujwe - kuva CRM na e-ubucuruzi kugeza kubisesengura no gukorana nitsinda - guhera ku $ 19 ku kwezi.

    Witegure gukora ubucuruzi bwawe bwuzuye bwuzuye?