Uhlolojikelele
I-Grouped-Query Attention (GQA) iyindlela yokunciphisa inkumbulo edingekayo phakathi nokukhiqizwa kombhalo ngokuvumela izinhloko zemibuzo ezimbalwa zabelane ngokhiye ofanayo namakhanda enani. Kwenza amamodeli amakhulu asheshe kakhulu ukuze asebenze cishe ngaphandle kokulahlekelwa kwekhwalithi.
Ukunakwa Kombuzo Ohlanganisiwe kuyingxenye yesitaki solimi-AI esisetshenziselwa ukufunda, ukukhiqiza, ukuhlukanisa, nokuguqula umbhalo nenkulumo ngesilinganiso.
I-Deep Dive
Kusendlalelo sokunaka esinamakhanda amaningi, inhloko ngayinye inemibuzo yayo, okhiye, namanani. Ngesikhathi sokukhiqiza, okhiye namanani awo wonke amathokheni adlule afakwe kunqolobane ('inqolobane ye-KV') ukuze imodeli ingawabuyiseli. Ngamakhanda amaningi nezimo ezinde, le cache iba yinkulu futhi ibusa umkhawulokudonsa wememori ngesikhathi sokunquma. I-GQA, eyethulwe Google abacwaningi ngo-2023, ihlanganisa izinhloko zemibuzo futhi inikeze iqembu ngalinye isethi eyodwa eyabelwe yokhiye namakhanda yenani. Uma unamakhanda emibuzo angu-32 kodwa amaqembu angu-8 KV kuphela, inqolobane ye-KV incipha cishe ngokuphindwe kane. Lokhu kuhlala phakathi kokunaka okugcwele kwamakhanda amaningi (ikhanda ngalinye lihlukene) kanye nokunakwa kwemibuzo eminingi (i-KV eyodwa eyabelwe kuwo wonke amakhanda), ithwebula iningi lejubane le-MQA kuyilapho igcina ikhwalithi iseduze nokunakwa okugcwele. I-Llama 2 70B kanye namamodeli amaningi akamuva ayamukela.
I-Technical Insight
Ikhwalithi yokunaka incike kakhulu ekubeni nezikhombisi-ndlela eziningi ezihlukene zemibuzo, kodwa ibekezelela ukwabelana ngokhiye namanani. I-GQA isebenzisa le-asymmetry: igcina yonke imibuzo kodwa iphindaphinda ikhanda ngalinye le-KV elabelwe kuyo yonke imibuzo eseqenjini layo. Ukongiwa kuza ngokulinganisa, lapho inqolobane ye-KV ingumthengi oyinhloko womkhawulokudonsa wenkumbulo; amakhanda e-KV ambalwa asho idatha encane okufanele ifundwe ngethokheni ngayinye ekhiqiziwe. Amamodeli avame 'ukuqeqeshwa' kafushane ukuze aguqule indawo yokuhlola enamakhanda amaningi abe yi-GQA.
Ukuphatha Ukunakwa Kwemibuzo Ehlanganisiwe
I-Grouped-Query Attention (GQA) iyindlela yokunciphisa inkumbulo edingekayo phakathi nokukhiqizwa kombhalo ngokuvumela izinhloko zemibuzo ezimbalwa zabelane ngokhiye ofanayo namakhanda enani. Kwenza amamodeli amakhulu asheshe kakhulu ukuze asebenze cishe ngaphandle kokulahlekelwa kwekhwalithi. Ukunakwa Kombuzo Ohlanganisiwe kuyingxenye yesitaki solimi-AI esisetshenziselwa ukufunda, ukukhiqiza, ukuhlukanisa, nokuguqula umbhalo nenkulumo ngesilinganiso. Ukuze wakhe ukuqonda okujulile, phatha Ukunakwa Kombuzo Ohlanganisiwe njengemodeli yokusebenza, hhayi isici esisodwa: chaza imiphumela efiselekayo, ucacise ukucabanga, futhi uhlukanise lokho isistimu engakwenza ngokwethembeka kulokho okusadinga ukwahlulela kochwepheshe.
Empeleni, amaqembu aqinile asebenzisa imiyalo yokuklama ye-Grouped-Query Attention, ukubuyisa, nokubuyekeza ama-loops njengohlelo olulodwa lokuxhumana oludidiyelwe. Babhala imibandela yempumelelo ecacile, ukuhlola okuqhathaniswa nedatha engokoqobo nokugeleza komsebenzi, futhi baphindaphinde ngokusekelwe kumaphethini okuhluleka aqashiwe esikhundleni sokuwina kwebhentshimakhi yesikhathi esisodwa. Yilapho ukuqonda kwethiyori kuguquka kube amandla ahlala njalo kuwo wonke umkhiqizo, inqubomgomo, kanye nokusebenza.
Ukugeleza komsebenzi wolimi kungahamba ngokushesha ngaphandle kokudela ukuvumelana. Ngesikhathi esifanayo, amaqiniso Akhohliwe angafaka imibiko buthule, ukugeleza kosekelo, noma imiphumela yocwaningo. Indlela eqine kakhulu iwukuhlanganisa isivinini sokuhlola nesiyalo sokuphatha: qhuba abashayeli bezindiza, bamba ubufakazi, ushicilele amalogi ezinqumo, futhi ubuyekeze izivikelo ngokuqhubekayo njengoba imodeli yokuziphatha, okulindelwe ngabasebenzisi, kanye nezimfuneko zokulawula zishintsha.
I-Strategic Impact
Ukugeleza komsebenzi wolimi kungahamba ngokushesha ngaphandle kokudela ukuvumelana.
Ukugeleza komsebenzi wolimi kungahamba ngokushesha ngaphandle kokudela ukuvumelana. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.
Yandisa ukufinyelela kuzo zonke izilimi nezitayela zokuxhumana.
Yandisa ukufinyelela kuzo zonke izilimi nezitayela zokuxhumana. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.
Amaqembu angachitha isikhathi esiningi ekwahluleleni kuyilapho i-automation isingatha impinda.
Amaqembu angachitha isikhathi esiningi ekwahluleleni kuyilapho i-automation isingatha impinda. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.
Ukuqaliswa Komhlaba Wangempela
I-Llama 2 70B kanye ne-Llama 3 isebenzisa i-GQA ukuze inikeze okuqukethwe okude ngenqolobane encane ye-KV
Kwehliswa inkumbulo ye-GPU ukuze imodeli yengxoxo enkulu ilingane kuma-accelerator ambalwa noma ashibhile
Ukusheshisa ukukhiqizwa kwamathokheni ngamathokheni kuma-API okukhiqiza lapho umkhawulokudonsa we-KV-cache uyibhodlela
Inika amandla osayizi benqwaba abakhulu bokusebenzela abasebenzisi abaningi ngesikhathi esisodwa ngaphandle kokuqeda inkumbulo
Amaphethini Okusebenzisa
Ukunakwa Kwemibuzo Ehlanganisiwe kuyasebenza
I-Llama 2 70B kanye ne-Llama 3 isebenzisa i-GQA ukuze inikeze okuqukethwe okude ngenqolobane encane ye-KV.
I-Llama 2 70B kanye ne-Llama 3 isebenzisa i-GQA ukuze inikeze okuqukethwe okude ngenqolobane encane ye-KV Amaqembu ngokuvamile athola imiphumela engcono uma echaza izinga eliphezulu ngaphambili, egcina indlela yokukhuphuka kwabantu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
Ukunakwa Kwemibuzo Ehlanganisiwe kuyasebenza
Kwehliswa inkumbulo ye-GPU ukuze imodeli yengxoxo enkulu ilingane kuma-accelerator ambalwa noma ashibhile.
Ukunciphisa inkumbulo ye-GPU ukuze imodeli yengxoxo enkulu ilingane kuma-accelerator ambalwa noma ashibhile Amaqembu ngokuvamile athola imiphumela engcono uma echaza izinga eliphezulu ngaphambili, agcine indlela yokukhuphuka yabantu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
Ukunakwa Kwemibuzo Ehlanganisiwe kuyasebenza
Ukusheshisa ukukhiqizwa kwamathokheni ngamathokheni kuma-API okukhiqiza lapho umkhawulokudonsa we-KV-cache uyibhodlela.
Ukusheshisa ukukhiqizwa kwamathokheni ngamathokheni kuma-API okukhiqiza lapho umkhawulokudonsa we-KV-cache uyibhodlela le-Bottleneck Amaqembu ngokuvamile athola imiphumela engcono uma echaza imingcele yekhwalithi ngaphambili, egcina indlela yokukhuphuka komuntu ngamacala asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
Ukunakwa Kwemibuzo Ehlanganisiwe kuyasebenza
Inika amandla osayizi benqwaba abakhulu bokusebenzela abasebenzisi abaningi ngesikhathi esisodwa ngaphandle kokuqeda inkumbulo.
Ukunika amandla osayizi benqwaba abakhulu bokusebenzela abasebenzisi abaningi ngesikhathi esisodwa ngaphandle kokuqeda inkumbulo Amaqembu ngokuvamile athola imiphumela engcono uma echaza izilinganiso zekhwalithi ngaphambili, egcina indlela yokukhuphuka yomuntu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
Izingozi & Guardrails
Amaqiniso akhonjiwe angafaka ngokuthula imibiko, ukugeleza kosekelo, noma imiphumela yocwaningo.
Ukuzwela okusheshayo kungadala imiphumela engahambisani kuzo zonke izicelo ezifanayo.
Idatha yombhalo ebucayi ingase idalulwe uma izilawuli zokufinyelela zibuthakathaka.
Ukuqalisa Umhlahlandlela
Chaza ifomethi yokuphumayo, ithoni, namazinga wekhwalithi ngaphambi kokukhishwa.
Chaza ifomethi yokuphumayo, ithoni, namazinga wekhwalithi ngaphambi kokukhishwa. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Izimpendulo eziyisisekelo ngemithombo ethembekile noma nini lapho ukunemba kubalulekile.
Izimpendulo eziyisisekelo ngemithombo ethembekile noma nini lapho ukunemba kubalulekile. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Gcina indawo yokuhlola isibuyekezo somuntu ukuze uthole imiphumela ephezulu.
Gcina indawo yokuhlola isibuyekezo somuntu ukuze uthole imiphumela ephezulu. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Landela amaphethini okuhluleka futhi uqeqeshe kabusha imiyalo noma ukuhamba komsebenzi njalo.
Landela amaphethini okuhluleka futhi uqeqeshe kabusha imiyalo noma ukuhamba komsebenzi njalo. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.