Hy-Nomas Uyasebenza

I-Hy-Nomas iyimodeli yomhlaba wevidiyo esebenzisanayo eyakhishwa njenge-open source yithimba le-Tencent Hunyuan. Ikhiqiza amavidiyo ngendlela ye-segmented autoregressive, isekela ukushintsha ngokukhululekile phakathi kweziqondiso zombhalo noma ukulawulwa komzila wekhamera phakathi kwezigaba, futhi isebenzisa i-lifelong KV cache ukugcina ukuhambisana kwesikhathi eside. Ikhodi nezisindo ze-Hy-Nomas sezikhishwe njenge-open source, futhi inikeza inguqulo edistiliwe kanye nenguqulo ye-INT8 quantized. Ngemva kokuthuthukiswa, i-RTX 4090 eyodwa ingayisebenzisa endaweni.

Ukulawulwa kombhalo okuguquguqukayo: Isekela iziqondiso zombhalo zanoma yiluphi uhlobo ukuze kulawulwe izinto, isitayela, indawo nezenzo emhlabeni okhiqizwayo.
Ukulawulwa komzila wekhamera: Ingakhiqiza ividiyo ngokusekelwe kumzila wokunyakaza kwekhamera, futhi ingahlanganiswa noma ishintshaniswe ngokukhululekile neziqondiso zombhalo.
Ukukhiqiza okusebenzisanayo kwemijikelezo eminingi: Isixhumi esibonakalayo seWeb sivumela ukufakwa kwemiyalelo emisha emingceleni yezigaba, ukuze kukhiqizwe ngokuqhubekayo emijikelezweni eminingi.
Ukwandiswa kweziqondiso (Nomas-PE): Imodeli eku-inthanethi eyakhelwe ngaphakathi iguqula ngokuzenzakalelayo imiyalelo yomsebenzisi yolimi oluvamile ibe yiziqondiso zokukhiqiza ezihlelekile.
Ukufakwa endaweni kudivayisi yabathengi: I-INT8 quantization kanye nokuthuthukiswa kwe-CPU offload kwenza i-RTX 4090 eyodwa ikwazi ukwenza inference.
Ukulungisa kahle nge-LoRA: Isekela ukulungiswa kahle kusetshenziswa idatha yevidiyo yangokwezifiso, futhi ingahlanganiswa nemodeli edistiliwe ukuze isetshenziselwe inference.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo zamaphrojekthi e-AI open source
Isakhiwo se-segmented autoregressive: Ividiyo ende ikhiqizwa ngezigaba; isigaba ngasinye sincike esiqondisweni samanje noma esignali yekhamera, futhi kugcinwa i-lifelong KV cache eqhubeka kuzo zonke izigaba. Lokhu kugcina indawo, izinto nesitayela kuhambisana ngokuhamba kwesikhathi, kwenze kube nokukhiqizwa komhlaba wesikhathi eside.
Ukusheshiswa kwe-inference nge-distillation: Ukuqeqeshwa kwe-distillation kunciphisa izinyathelo ze-denoising zibe ngu-4/6/8. Ngokuhambisana ne-PixelUMM decoder eqeqeshwe kahle, kusingathwa ama-latent variables e-32-channel DCAE-V, kwehliswe kakhulu ukubambezeleka kwe-inference ngesikhathi kugcinwa ikhwalithi.
Indlela ye-inference esebenzisa inkumbulo encane yevidiyo: Izisindo ze-DiT ne-KV cache zigcinwa nge-INT8, kanti ukubala kugcinwa ku-BF16; umongo womlando usikwa nge-sliding window ukuze ungakhuli ngokungenamkhawulo; i-text encoder, i-DiT ne-VAE zishintshwa phakathi kwe-CPU ne-GPU ngokwezigaba, kwehliswe inkumbulo ephezulu yevidiyo ibe sezingeni elingakwazi ukuphathwa yikhadi lemifanekiso labathengi.
Ukuqeqeshwa okuphoqelelwe kwezigaba nge-LoRA: Ukulungiswa kahle kusebenzisa i-checkpoint engadistilwanga njengemodeli eyisisekelo, kusetshenziswe i-segment forcing training pipeline, futhi kuqeqeshwe kuphela amapharamitha e-LoRA (rank/alpha 256). Ngemva kokuqedwa kokuqeqeshwa, izisindo ze-LoRA zihlanganiswa futhi ku-checkpoint edistiliwe, kutholakale imodeli ye-inference egcina ukusheshiswa kwe-distillation futhi inamandla okwenziwa ngokwezifiso.

Ukushintsha ngokukhululekile phakathi kwezindlela ezimbili: Esigabeni ngasinye sokukhiqiza kungashintshwa ngokukhululekile phakathi kweziqondiso zombhalo nokulawulwa komzila wekhamera, okwenza ukusebenzisana kube nezimo.
Ukuhambisana kwesikhathi eside: I-lifelong KV cache iqinisekisa ukuqhubeka kwezinto, indawo nesitayela kuzo zonke izigaba, ivimbele ividiyo ende ukuthi iphambuke.
Iyasebenza kumadivayisi abathengi: I-INT8 quantization, i-sliding window kanye ne-CPU offload kwenza i-RTX 4090 eyodwa ikwazi ukwenza inference endaweni.
I-inference eshesha kakhulu: Ngemva kwe-distillation, kudingeka kuphela izinyathelo ze-denoising ezingu-4~8, kwehliswe kakhulu ukubambezeleka kokukhiqiza.
I-open source ngokuphelele: Ikhodi, izisindo, ukuqeqeshwa kanye nenqubo yokulebula idatha konke kuyatholakala emphakathini, futhi ilayisensi ye-Apache 2.0 ivumela ukusetshenziswa kwezohwebo.

Iwebhusayithi esemthethweni yephrojekthi: https://hy-nomas.hy.tencent.com/
Indawo yokugcina ye-GitHub: https://github.com/Tencent-Hunyuan/Hy-Nomas
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/tencent/Hy-Nomas

Ukuthuthukiswa kwama-prototype emidlalo ye-AI: Khiqiza ngokushesha imihlaba yemidlalo esebenzisanayo usebenzisa imiyalelo yombhalo nekhamera, ukuze kuhlolwe imiqondo yokudlala ngezindleko eziphansi.
Ukuqeqeshwa kokulingisa amarobhothi: Sebenzisa i-Robotics LoRA esemthethweni ukukhiqiza izindawo ezingokoqobo zokusebenza kwengalo yerobhothi, unikeze ubuhlakani obuhlanganisiwe idatha yokufundisa eyakhiwe.
Ukulungiselela amafilimu ne-storyboard: Sebenzisa i-LoRA enesitayela sefilimu ukukhiqiza izithombe ezinde ezihambisanayo, usize abaqondisi ekuklameni ukunyakaza kwekhamera nasekubukeni kusengaphambili.
Ezokuvakasha ezibonakalayo nokwethulwa kwezindawo: Sebenzisa ukulawulwa komzila wekhamera ukukhiqiza izindawo zezokuvakasha, amahholo emibukiso nezindawo zezindlu ezingahanjwa, esikhundleni sokumodela kwe-3D kwendabuko.
Ucwaningo nemfundo ye-AI: Ikhodi ephelele yokuqeqesha, yokulungisa kahle kanye nokulebula idatha iyi-open source, futhi ingasetshenziswa njengesisekelo sokufundisa nokucwaninga emkhakheni wamamodeli omhlaba.