ForgeStencil – uhlelo lwe-Stencil lokucwaninga nokuthunyelwa okuzenzakalelayo ngokuphelele oluqaliswe yi-Mianbi Intelligent Uyasebenza

I-ForgeStencil iwuhlelo lokuqala emhlabeni lokuthuthukisa olusebenzisa i-AI olwesekela ucwaningo oluzenzakalelayo nokuthunyelwa okuzenzakalelayo kwe-Stencil, olwethulwe ngokubambisana yi-Mianbi Intelligent nomphakathi ovulekile we-OpenBMB. Ngokusekelwe ekuvalekeni kwama-agent amabili, i-Kernel Agent ne-App Agent, lwenza yonke inqubo kusukela ekutholeni amasu okuthuthukisa kuya ekuhlanganiseni isofthiwe yangempela yokukhiqiza, ngaphandle kokungenelela komuntu. Esontweni elilodwa, uhlelo lwaqeda ukuthuthukiswa kusukela ekuqaleni kuye ekugcineni kwezinhlelo ezingaphezu kwe-100 zokubala zezimboni nezesayensi, lwafinyelela ekusheshiseni okumaphakathi okuphindwe ngo-1.41, futhi luhlanganisa imikhakha yezimboni engu-8 kanye nemikhakha yesayensi engu-5, okuhlanganisa ukuhlola uwoyela negesi, ukulingisa amandla kazibuthe kagesi, nezithombe zezokwelapha.

Ukuthuthukisa okuvaliwe kwama-Agent amabili: I-Kernel Agent icwaninga ngokuzenzakalelayo futhi ihlanganise ama-CUDA Kernel asebenza kahle, kanti i-App Agent ithola izindawo ezishisa kakhulu, yenza ukwakhiwa kwama-operator, iqinisekise ukunemba, futhi ihlanganise uhlelo.
Ukutholwa kwamasu okuzenzakalelayo: Ama-Agent ahlola ngokuzenzakalelayo amasu okuthuthukisa afana nokuhlukanisa ngamabhulokhi, ukuhlanganisa, ukuhlelwa, izinga lokusebenzisa izinsiza, nokwakhiwa kabusha kohlangothi lwe-host, kunokuba afune ngaphakathi kwendawo engashintshi.
Ukuthunyelwa kwezinhlelo zangempela: Kuthuthukiswa ngqo isofthiwe esezingeni lokukhiqiza, kuqinisekiswe ukusebenza kusukela ekuqaleni kuye ekugcineni kusetshenziswa ikhodi ye-GPU yohlelo uqobo njengesisekelo, hhayi i-Benchmark eyenziwe lula.
Iphrothokholi yokukala ehlolwayo: Ngokusebenzisa iswishi eyodwa yemvelo, izindlela zangempela nezithuthukisiwe zishintshaniswa, kusetshenziswa ukuhlolwa nokubala isikhathi okwakhelwe ohlelweni ukuze kuqinisekiswe ukuthi ukusheshisa kuyiqiniso futhi kunokwethenjelwa.
Ukuhambisana nama-GPU ezizukulwane ezahlukene: Isekela ukusatshalaliswa kwesakhiwo sesikhathi sokusebenza se-A100/H100/B200, futhi ikhodi efanayo ikhetha ngokuzenzakalelayo indlela ye-Kernel ehamba phambili ngokwesakhiwo.

Landela i-WeChat, uphendule ngokuthi “开源”, bese ujoyina iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Ukwakheka kokubambisana kwama-Agent amabili: Uhlelo lwakhiwe yi-Kernel Agent ne-App Agent olusebenza njengesiyingi esivaliwe. I-Kernel Agent iphethe ukutholwa kolwazi lwethiyori, isebenzisa umjikelezo we-Plan→Code→Profile ukuze icwaninge ngokuzenzakalelayo futhi ihlanganise ama-operator e-Stencil asondela emkhawulweni womzimba wehadiwe; i-App Agent iphethe ukusetshenziswa kwezobunjiniyela, ithole izindawo ezibambezela ukusebenza kohlelo, isungule isisekelo se-GPU, isebenzise isizindalwazi solwazi se-matrix yama-operator esakhiwe yi-Kernel Agent ukuze yakhe ukuthuthukiswa okuqondene nohlelo, bese iqinisekisa ukunemba futhi ihlanganise lokho nohlelo lwangempela.
Ukuthuthuka kwe-matrix yama-operator nesizindalwazi solwazi: Ngesikhathi sokuthuthukisa, i-Kernel Agent iyaqhubeka nokwakha i-matrix yama-operator ekhethekile njengesizindalwazi solwazi, ukuze i-App Agent ikwazi ukuphinda isebenzise ubuchwepheshe obukhona. Ama-Agent amaningi asebenza ngokuhambisana abelana ngalesi sizindalwazi, okuvumela ukwabelana kwesipiliyoni ngesikhathi sangempela nokuthuthuka okuhlangene okuvumelanayo, ngaleyo ndlela kweqe umkhawulo wokudlulisa ngobukhulu ulwazi lochwepheshe babantu.
Iphrothokholi ehlolwayo kusukela ekuqaleni kuye ekugcineni: Ukukala kusebenzisa kuphela uhlelo lwangempela lwe-GPU olusezingeni lokukhiqiza njengesisekelo. Indlela yangempela nethuthukisiwe yehlukaniswa kuphela ngeswishi eyodwa ye-USE_OURS, futhi indlela yangempela ihambisana ncamashi namabhayithi omthombo ophezulu. Kusetshenziswa ukuhlolwa kokunemba nezibali-sikhathi ezakhelwe ohlelweni; ngemva kokusebenza okushintshanayo kaningi, kuthathwa inani eliphakathi, bese kubikwa ukusheshisa okumaphakathi kwejiyomethri kwazo zonke izimo ezijwayelekile, okwenza kube nzima kakhulu ukukhohlisa ezingeni lohlelo.
Imodi ye-Patch nokungabopheleli kwekhodi yomthombo: Akukho khodi yomthombo yomuntu wesithathu ebophelelwayo. Uhlelo ngalunye lunikeza kuphela irekhodi lomthombo ophezulu, iskripthi sokuthola ikhodi, kanye ne-patch yokuhlanganisa. I-vendor.sh idonsa ngokuzenzakalelayo inguqulo ethile yekhodi yomthombo ephezulu, iqinisekise ukuhlanzeka kwamalayisensi nokuphindaphindeka kwemiphumela.

Ukulungiselela imvelo: Klona i-repository futhi uqinisekise ukuthi une-NVIDIA GPU (kuqinisekiswe nge-A100), i-CUDA 12.x, i-compiler ye-C++17, kanye nemvelo ye-Python 3.9+.
Ukuhlola i-operator ngokushesha: Sebenzisa python tools/run.py --stencil star_1 --shape 256 --gpu 0 ukuze ulinganise ukusebenza kwe-operator eyodwa kungakapheli iminithi elingu-1 futhi ukuqhathanise nesisekelo se-Halide.
Ukuphinda ukwakhe konke kusukela ekuqaleni kuye ekugcineni: Ngena kufolda yohlelo oluqondisiwe, usebenzise ./vendor.sh ukuze udonse ikhodi yomthombo ephezulu, usebenzise i-patch yokuhlanganisa, bese usebenzisa python ../../harness/run_e2e.py --app --gpu auto ukuze uqinisekise uhlelo lwangempela.
Ukuqhuba umsebenzi wama-Agent: Bheka ku-agents/README.md ukuze ulungiselele futhi uqalise umjikelezo wokuthuthukisa okuzenzakalelayo we-Kernel Agent ne-App Agent.
Ukuhlola irejista yemiphumela: Bheka idatha yokusheshisa ehloliwe yezinhlelo eziyi-100 eziqinisekisiwe ku-results/integration_registry.json.

Akukho kungenelela komuntu: Yonke inqubo kusukela ekuhlaziyeni uhlelo, ekutholeni izindawo ezishisa kakhulu, ekwakheni i-Kernel, kuya ekuqinisekiseni ukuhlanganiswa yenziwa ngokuzenzakalelayo, ngaphandle kokudinga ochwepheshe be-HPC ukuthi benze izinqumo.
Igxile ezimweni zangempela: Kuthuthukiswa ngqo isofthiwe yezimboni esezingeni lokukhiqiza; u-42% uyisofthiwe yezimboni esetshenziswa ekukhiqizeni kwangempela, futhi isisekelo yikhodi ye-GPU yohlelo uqobo.
Ukuthuthuka okukhulu kokusebenza kahle kocwaningo: Ukuthuthukisa uhlelo olulodwa kwehliswa kusuka ezinyangeni eziningi kuye kungakapheli usuku olulodwa, izinhlelo ezingu-100+ ziqedwa ngesonto elilodwa, okulingana nokonga cishe iminyaka engu-20–30 yomsebenzi wokuthuthukisa.
Ukusebenza okunamandla: Ukusheshisa okumaphakathi kusukela ekuqaleni kuye ekugcineni kuphindwe ngo-1.41; izinhlelo ezingu-43% zifinyelela okungenani ku-1.5; isilinganiso sejiyomethri ezingeni lama-operator singu-2.16, sihamba phambili kunesisekelo esihle kakhulu esivulekile.
Ukusimama kuzo zonke izakhiwo: Isekela ukusatshalaliswa kwesikhathi sokusebenza kwama-NVIDIA GPU ezizukulwaneni eziningi, futhi lapho isakhiwo sithuthukiswa, ingaphinda yakhe ngokuzenzakalelayo i-Kernel ukuze ithole izici zehadiwe entsha.

I-GitHub repository: https://github.com/OpenBMB/ForgeStencil

Ukuhlola kokuzamazama komhlaba kukawoyela negesi: Kuthuthukiswa ama-algorithm abalulekile afana ne-RTM reverse-time migration kanye ne-mini-app ye-Total, kusheshiswe inqubo yokucubungula idatha yokuhlola uwoyela.
Ukwakha ukulingisa amandla kazibuthe kagesi: Kuthuthukiswa ukusebenza kokuxazulula izibalo ze-Maxwell kugridi ye-Yee, njengaku-gprMax/FDTD, ukuze kusetshenziswe ekuhlaziyeni amandla kazibuthe kagesi kuma-radar nama-chip.
Ukwakhiwa kabusha kwezithombe zezokwelapha: Kusheshiswa imithwalo yokubala yezithombe zezokwelapha, okuhlanganisa ukwakhiwa kabusha kwe-MRI engeyona eye-Cartesian kanye ne-backprojection ye-digital breast tomosynthesis.
Isimo sezulu ne-astrophysics: Kuthuthukiswa izinhlelo zokubala zesayensi ezisebenzisa kakhulu ama-Stencil, njengamandla omkhathi nokulingisa i-cosmology.
Ukubala kwezimali ze-quantitative: Kuhlinzekwa ukuthuthukiswa okuzenzakalelayo kwezindawo ezishisa kakhulu ze-Stencil kumalabhulali entengo ezikhungo zezimali afana ne-QuantLib, kwehliswe ukubambezeleka kokubala kokuhweba.