I-Google isimemezele ukushicilelwa kwe-Gemma, imodeli yolimi lomshini olukhulu olwakhiwe kusetshenziswa ubuchwepheshe obusetshenziselwa ukwakha imodeli ye-Gemini chatbot, ezama ukuncintisana ne-ChatGPT. Imodeli itholakala ngezinhlobo ezine, ezihlanganisa amapharamitha ayizigidi eziyizinkulungwane ezimbili nezingu-2, ekubukweni okuyisisekelo kanye nengxoxo elungiselelwe. Izinketho ezinamapharamitha ayizigidi eziyizinkulungwane ezingu-7 zifanele ukusetshenziswa kuzinhlelo zokusebenza zabathengi futhi zine-CPU eyanele ukuzicubungula. Izinketho ezinamapharamitha ayizigidi eziyizinkulungwane ezingu-2 zidinga ihadiwe enamandla kakhulu kanye ne-GPU noma i-TPU.
Phakathi kwezindawo zokusetshenziswa kwemodeli ye-Gemma ukwakhiwa kwezinhlelo zezingxoxo nabasizi abangokoqobo, ukukhiqizwa kombhalo, ukukhiqizwa kwezimpendulo zemibuzo ebuzwa ngolimi lwemvelo, isifinyezo kanye nokwabiwa kokuqukethwe, incazelo yengqikithi yemigomo nemibandela, ukulungiswa kwamaphutha. embhalweni, usizo ekufundeni izilimi. Isekela ukwakhiwa kwezinhlobo ezahlukene zedatha yombhalo, okuhlanganisa izinkondlo, ikhodi ngezilimi zokuhlela, ukubhala kabusha imisebenzi ngamanye amagama, nokukhiqiza izinhlamvu kusetshenziswa isifanekiso. Ngesikhathi esifanayo, imodeli inesayizi encane, evumela ukuthi isetshenziswe kumishini yakho enezinsiza ezilinganiselwe, isibonelo, kuma-laptops ajwayelekile nama-PC.
Ilayisensi eyimodeli ivumela ukusetshenziswa kwamahhala nokusabalalisa hhayi kuphela ocwaningweni namaphrojekthi womuntu siqu, kodwa nasemikhiqizweni yokuhweba. Ukudalwa nokushicilelwa kwezinguqulo eziguquliwe zemodeli nakho kuvunyelwe. Ngesikhathi esifanayo, imigomo yokusetshenziswa iyakwenqabela ukusetshenziswa kwemodeli ukwenza izenzo ezinonya futhi idinga, noma nini lapho kunokwenzeka, ukusebenzisa inguqulo yakamuva ye-Gemma emikhiqizweni yakho.
Usekelo lokusebenza namamodeli we-Gemma seluvele lwengeziwe kukhithi yamathuluzi ye-Transformers kanye ne-Responsible Generative AI Toolkit. Ukuze uthuthukise imodeli, ungasebenzisa uhlaka lwe-Keras nezingemuva ze-TensorFlow, i-JAX ne-PyTorch. Kungenzeka futhi ukusebenzisa i-Gemma ngezinhlaka ze-MaxText, i-NVIDIA NeMo kanye ne-TensorRT-LLM.
Ubukhulu bomongo obucatshangelwe yimodeli ye-Gemma ngamathokheni ayizinkulungwane ezingu-8 (inani lamathokheni imodeli engakwazi ukucubungula nokukhumbula lapho ikhiqiza umbhalo). Ukuze uqhathanise, usayizi womongo wamamodeli we-Gemini ne-GPT-4 ungamathokheni ayizinkulungwane ezingu-32, kanti imodeli ye-GPT-4 Turbo iyizinkulungwane ezingu-128. Imodeli isekela isiNgisi kuphela. Mayelana nokusebenza, imodeli ye-Gemma-7B ingaphansi kancane kumodeli ye-LLama 2 70B Chat futhi ingaphambili kancane kunezinhlobo ze-DeciLM-7B, PHI-2 (2.7B) ne-Mistral-7B-v0.1. Ekuqhathanisweni kwe-Google, imodeli ye-Gemma-7B ingaphambili kancane kune-LLama 2 7B/13B ne-Mistral-7B.
Source: opennet.ru