Uhlolojikelele
I-SwiGLU iwumsebenzi wokwenza kusebenze onesango ophindaphinda ukuqagela okukodwa komugqa kokufakwayo ngokuqagela kwesibili okusetshenziselwe i-Swish, okusebenza njengesango elifundekayo, elincike kudatha ngaphakathi kwezendlalelo zokudlulisela phambili okuphakelayo kwe-transformer. Ihlala ithuthukisa ikhwalithi yemodeli yolimi, yingakho cishe yonke i-LLM yesimanje iyisebenzisa.
I-SwiGLU kanye ne-Gated Activations iyibhulokhi yokwakha yobuchwepheshe ethinta ikhwalithi yemodeli, izindleko zengqalasizinda, ukubambezeleka, nokuthembeka esikalini.
I-Deep Dive
Ibhulokhi yokuphakela phambili kwe-transformer ejwayelekile iyizendlalelo ezimbili eziqondile ezine-ReLU noma i-GELU phakathi. I-Gated Linear Units, ehlongozwe ngu-Dauphin et al. ngo-2016, hlukanisa i-projection yokuqala ibe izingxenye ezimbili futhi usebenzise ingxenye eyodwa ukuze isango enye ngokuphindaphinda okuhlakaniphile. I-SwiGLU, eyaduma nguNoam Shazeer ngo-2020, isebenzisa umsebenzi we-Swish (SiLU) walelo sango: okukhiphayo = (Swish(xW) * (xV)) W2, enamatrices amathathu esisindo esikhundleni samabili. Isango lokungena livumela inethiwekhi ngokukhetha ukuthi idlule noma icindezele ulwazi ngosayizi ngamunye. Ngenxa yokuthi ukungeza i-matrix yesithathu kukhulisa amapharamitha, ukusetshenziswa kufinyeza ubukhulu obufihliwe ukuze bucishe kube izingxenye ezimbili kwezintathu ukuze ikhompuyutha ephelele ihlale iqhathaniswa ne-GELU MLP. Ukuhlola kuka-Shazeer kubonise izinzuzo ezilinganisekayo zokudideka, futhi i-LLaMA, i-PaLM, ne-Mistral zonke zakwamukela.
I-Technical Insight
I-Swish ithi x * sigmoid(beta*x), umsebenzi obushelelezi, ongewona we-monotonic, ngokungafani ne-ReLU, ovumela amanani amancane angalungile ngawo. Ku-SwiGLU igatsha 'lesango' i-Swish(xW) ikhiqiza amanani aseduze no-0 noma 1 aphindaphinda ingxenye 'yenani' yegatsha xV, ngakho umnikelo weyunithi ngayinye efihliwe ulungiswa ngesignali efundiwe, encike kokufakiwe. I-matrix yesisindo sesithathu yizindleko; iqhinga lokubili kokuthathu losayizi ofihliwe ligcina ibhajethi ye-FLOP ifaniswe nesendlalelo sokuphakela phambili kwe-vanilla.
I-Mastering SwiGLU kanye ne-Gated Activations
I-SwiGLU iwumsebenzi wokwenza kusebenze onesango ophindaphinda ukuqagela okukodwa komugqa kokufakwayo ngokuqagela kwesibili okusetshenziselwe i-Swish, okusebenza njengesango elifundekayo, elincike kudatha ngaphakathi kwezendlalelo zokudlulisela phambili okuphakelayo kwe-transformer. Ihlala ithuthukisa ikhwalithi yemodeli yolimi, yingakho cishe yonke i-LLM yesimanje iyisebenzisa. I-SwiGLU kanye ne-Gated Activations iyibhulokhi yokwakha yobuchwepheshe ethinta ikhwalithi yemodeli, izindleko zengqalasizinda, ukubambezeleka, nokuthembeka esikalini. Ukuze wakhe ukuqonda okujulile, phatha i-SwiGLU kanye ne-Gated Activations njengemodeli yokusebenza, hhayi isici esisodwa: chaza imiphumela efiselekayo, ucacise ukucabanga, futhi uhlukanise lokho uhlelo olungakwenza ngokwethembeka kulokho okusadinga ukwahlulela kochwepheshe.
Empeleni, amaqembu aqinile asebenzisa i-SwiGLU kanye ne-Gated Activations athuthukisa izakhiwo, idatha, nokukhetha kwengqalasizinda ngokumelene nokuthembeka nezindleko. 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.
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka. Ngesikhathi esifanayo, Ukuthuthukisa ibhentshimakhi eyodwa kungafihla ubuthakathaka obubanzi besistimu. 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
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka.
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.
Imfundo yobuchwepheshe isiza amaqembu ukuthi akhethe isitaki esifanele, hhayi nje esisha.
Imfundo yobuchwepheshe isiza amaqembu ukuthi akhethe isitaki esifanele, hhayi nje esisha. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.
Izinketho ezingcono zobunjiniyela zinciphisa izehlakalo ezinokwethenjelwa ekukhiqizeni.
Izinketho ezingcono zobunjiniyela zinciphisa izehlakalo ezinokwethenjelwa ekukhiqizeni. 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, i-PaLM, ne-Mistral ibuyisela isendlalelo sokuphakela phambili se-GELU nge-SwiGLU ukuze kwehle ukudideka ngekhompyutha elinganayo.
Ubukhulu obufihliwe bulinganiswa cishe kuzingxenye ezimbili kwezintathu (8/3 d) ukuze i-matrix eyengeziwe yokufaka isango ingavutheli ama-FLOP
Ingxubevange yamamodeli afana ne-Mixtral asebenzisa i-SwiGLU ivimba njengenethiwekhi yokuphakela phambili kochwepheshe
Umbono kanye nama-multimodal transformer aboleka i-GeGLU/SwiGLU gating ukuze bathuthukise ama-sublayers abo e-MLP
Amaphethini Okusebenzisa
I-SwiGLU kanye ne-Gated Activations ekusebenzeni
I-LLaMA, i-PaLM, ne-Mistral zishintsha isendlalelo sokuphakela phambili kwe-GELU nge-SwiGLU ukuze kwehle ukudideka ngekhompyutha elinganayo.
I-LLaMA, i-PaLM, ne-Mistral ithatha indawo ye-GELU feed-forward layer nge-SwiGLU ukuze yehlise ukudideka ekulinganisweni kwe-Compute Teams ngokuvamile athola imiphumela engcono uma echaza imingcele yekhwalithi ngaphambili, egcina indlela yokukhuphuka komuntu ngamacala asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
I-SwiGLU kanye ne-Gated Activations ekusebenzeni
Ubukhulu obufihliwe bulinganiswa cishe kuzingxenye ezimbili kwezintathu (8/3 d) ukuze i-matrix eyengeziwe yokufaka isango ingavutheli ama-FLOP.
Ubukhulu obufihliwe bulinganiswa cishe kuzingxenye ezimbili kwezintathu (8/3 d) ukuze i-matrix eyengeziwe yokufaka isango ingavutheli ama-FLOPs Amaqembu ngokuvamile athola imiphumela engcono uma echaza imingcele yekhwalithi ngaphambili, agcina indlela yokukhuphuka komuntu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
I-SwiGLU kanye ne-Gated Activations ekusebenzeni
Ingxubevange yamamodeli afana ne-Mixtral asebenzisa i-SwiGLU ivimba njengenethiwekhi yokuphakela phambili kochwepheshe ngamunye.
Ingxubevange yamamodeli afana nokusetshenziswa kwe-Mixtral i-SwiGLU ivimba njengenethiwekhi yenethiwekhi yochwepheshe ngamunye Amathimba ngokuvamile athola imiphumela engcono lapho echaza imingcele yekhwalithi ngaphambili, egcina indlela yokukhuphuka komuntu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
I-SwiGLU kanye ne-Gated Activations ekusebenzeni
Umbono kanye nama-multimodal transformer aboleka i-GeGLU/SwiGLU isango ukuze bathuthukise ama-sublayers abo e-MLP.
Umbono kanye nama-multimodal transformer aboleka i-GeGLU/SwiGLU gating ukuze athuthukise Amaqembu abo angaphansi kwe-MLP Amaqembu ngokuvamile athola imiphumela engcono lapho echaza imingcele yekhwalithi ngaphambili, egcina indlela yokukhuphuka kwabantu yamacala asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.
Izingozi & Guardrails
Ukuthuthukisa ibhentshimakhi eyodwa kungafihla ubuthakathaka obubanzi besistimu.
Izindleko zengqalasizinda nezokulungisa zivame ukubukelwa phansi.
Izikhala zokuphepha nokubonakala zingakhula njengoba izinhlelo ziba nzima kakhulu.
Ukuqalisa Umhlahlandlela
Chaza ukubambezeleka, ikhwalithi, nezindleko ezihlosiwe ngaphambi kokuqaliswa.
Chaza ukubambezeleka, ikhwalithi, nezindleko ezihlosiwe ngaphambi kokuqaliswa. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Ibhentshimakhi ngaphansi komthwalo wangempela nezimo zedatha.
Ibhentshimakhi ngaphansi komthwalo wangempela nezimo zedatha. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Ukuqapha amathuluzi amaphutha, ukukhukhuleka, nomthelela wabasebenzisi.
Ukuqapha amathuluzi amaphutha, ukukhukhuleka, nomthelela wabasebenzisi. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.
Lungiselela izindlela zokuhlehlisa nezigameko ngaphambi kokukala.
Lungiselela izindlela zokuhlehlisa nezigameko ngaphambi kokukala. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.