Ubuyobozi bwa tekiniki

Icyitonderwa

Flash Attention nuburyo bwubwenge bwo kubara intambwe yimbere muri Transformers utarigeze wandika matrix nini yo kwitondera buhoro buhoro.

Incamake

Flash Attention nuburyo bwubwenge bwo kubara intambwe yimbere muri Transformers utarigeze wandika matrix nini yo kwitondera buhoro buhoro. Cyakora imiterere-ndende-moderi yihuta cyane kandi yibuka-neza idahinduye imibare.

Flash Attention ni tekiniki yubaka tekinike igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo.

Kwibira cyane

Icyitonderwa gisanzwe kigereranya buri kimenyetso nikindi kimenyetso, gitanga amanota ya N-by-N ikura inshuro enye hamwe nuburebure bwikurikiranya. Na Naive, iyo matrix yandikiwe kandi isomwe uhereye muri GPU yibuka ryinshi rya HBM (HBM), kandi ko guhinduranya - ntabwo ari kugwira - nicyo cyuho. Flash Attention, yatangijwe na Tri Dao na bagenzi be muri 2022, itunganya ibarura kugirango matrix itabikwa neza. Itunganya ibibazo, urufunguzo, nagaciro mumashusho mato ahuye byihuse kuri chip SRAM, ibara ibisubizo byigice, kandi ikabidoda hamwe ikoresheje interineti ikora-softmax trick. Ibisohoka ni imibare isa nibisanzwe ariko ikoresha umurongo wibuke kandi ikora inshuro nyinshi byihuse, cyane cyane kumurongo muremure.

Ubushishozi

Amayeri yingenzi ni tiling wongeyeho softmax kumurongo. Ubusanzwe Softmax ikenera umurongo wose wamanota kugirango ibare icyerekezo cyayo, ariko Flash Attention ikomeza gukora ntarengwa kandi ikora amafaranga uko itemba kuri tile, ikuraho ibyasohotse mbere igice kuburyo ibisubizo byanyuma aribyo. Kuberako amanota aringaniye aguma muri SRAM (ordre yubunini bwihuse kurenza HBM), algorithm irabizi IO: igabanya kwibuka gusoma no kwandika aho gukora imibare mibare.

Kumenya Flash

Flash Attention nuburyo bwubwenge bwo kubara intambwe yimbere muri Transformers utarigeze wandika matrix nini yo kwitondera buhoro buhoro. Cyakora imiterere-ndende-moderi yihuta cyane kandi yibuka-neza idahinduye imibare. Flash Attention ni tekiniki yubaka tekinike igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo. Kugirango wubake byimbitse, fata Flash Attention nkicyitegererezo gikora, ntabwo ari ikintu kimwe: gusobanura ibyagezweho, gusobanura ibitekerezo, no gutandukanya ibyo sisitemu ishobora gukora byizewe nibisaba guca imanza zinzobere.

Mu myitozo, amakipe akomeye akoresha Flash Attention atezimbere ubwubatsi, amakuru, hamwe nibikorwa remezo birwanya kwizerwa nigiciro. Bandika ibipimo ngenderwaho byerekana intsinzi, bagerageza kurwanya amakuru afatika hamwe nakazi keza, kandi bagasubiramo bashingiye kubikorwa byagaragaye ko batsinzwe aho gutsinda inshuro imwe. Aha niho imyumvire yubumenyi ihinduka mubushobozi burambye kubicuruzwa, politiki, nibikorwa.

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka. Mugihe kimwe, Guhindura igipimo kimwe gishobora guhisha intege nke za sisitemu. Uburyo bukomeye cyane ni uguhuza umuvuduko wikigereranyo hamwe na disipuline yimiyoborere: kuyobora abaderevu, gufata ibimenyetso, gutangaza ibyemezo byicyemezo, no gukomeza kuvugurura uburyo bwo kwirinda nkimyitwarire yicyitegererezo, ibyo abakoresha bategereje, nibisabwa n'amategeko bigenda bihinduka.

Ingaruka z'Ingamba

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.

Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.

Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.

Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa. Mubikorwa byujuje ubuziranenge, ibi bihindurwa mumategeko agenga imikorere, imipaka nyirubwite, hamwe n'imihango yo gusubiramo kenshi kugirango amakipe ashobore kwigirira ikizere aho gupima ibidasobanutse.

Ejo hazaza ha Flash

Flash Attention yahindutse inyubako idasanzwe, hamwe na FlashAttention-2 na FlashAttention-3 ikanda byinshi byinjira muri GPU nshya nka H100 mugutezimbere igabana ryakazi no gukoresha inzira nke za FP8. Witege ko uzakomeza gushushanya hamwe nibikoresho, guhuza imbaraga mumahugurwa no gufata umwanzuro, hamwe nibihinduka byahujwe gake, kunyerera-idirishya, hamwe-birebire-by-ibitekerezo. Nkuko imiterere ya Windows irambuye kuri miriyoni yikimenyetso, IO-izi intangiriro nkiyi ikomeza kuba ingenzi kugirango ukomeze kwibuka no kwihuta mubikorwa.

Gushyira mu bikorwa Isi

Guhugura imiterere yindimi nini nka Llama na GPT-urwego rwa sisitemu hamwe na Windows ndende ndende ku giciro cyo kwibuka gito.

Gukorera abafasha kuganira byihuse mukwihutisha icyiciro cyambere aho isomwa rirerire ryambere.

Gushoboza inyandiko-isesengura ibikoresho byinjiza ibitabo byose cyangwa codebase mugukora ibitekerezo birebire byashoboka kuri GPU imwe.

Imbaraga zicyerekezo hamwe namajwi Guhindura amajwi aho ibisubizo bihanitse byinjiza birema ibimenyetso birebire cyane.

Uburyo bwo Gushyira mu bikorwa

Flash Itondere mubikorwa

Guhugura imiterere yindimi nini nka Llama na GPT-urwego rwa sisitemu hamwe na Windows ndende ndende ku giciro cyo kwibuka gito.

Guhugura imvugo nini yindimi nka Llama na GPT-urwego rufite urwego rurerure rwa Windows ku giciro cyo kwibuka cyo hasi Amakipe ubusanzwe abona ibisubizo byiza iyo asobanuye ibipimo ngenderwaho byimbere, agakomeza inzira yo kuzamura abantu kubibazo, kandi akurikirana inyungu zibyara umusaruro hamwe nibiciro byamakosa mugihe runaka.

Flash Itondere mubikorwa

Gukorera abafasha kuganira byihuse mukwihutisha icyiciro cyambere aho isomwa rirerire ryambere.

Gukorera abafasha kuganira byihuse mukwihutisha icyiciro cyambere aho isomwa rirerire ryambere risomwa Amakipe mubisanzwe abona ibisubizo byiza iyo asobanuye ibipimo byujuje ubuziranenge imbere, agakomeza inzira yo kuzamura abantu kubibazo, kandi akurikirana inyungu zibyara umusaruro hamwe nibiciro byamakosa mugihe runaka.

Flash Itondere mubikorwa

Gushoboza inyandiko-isesengura ibikoresho byinjiza ibitabo byose cyangwa codebase mugukora ibitekerezo birebire byashoboka kuri GPU imwe.

Gushoboza inyandiko-isesengura ibikoresho byinjira mubitabo byose cyangwa codebase mugukurikirana ibitekerezo birebire bishoboka mumakipe amwe ya GPU mubisanzwe abona ibisubizo byiza mugihe basobanuye ibipimo byimbere imbere, bagakomeza inzira yo kuzamuka kwabantu kubibazo byimbere, kandi bagakurikirana inyungu zibyara umusaruro nibiciro byamakosa mugihe runaka.

Flash Itondere mubikorwa

Imbaraga zicyerekezo hamwe namajwi Guhindura amajwi aho ibisubizo bihanitse byinjiza birema ibimenyetso birebire cyane.

Guha imbaraga iyerekwa hamwe n'amajwi ahinduranya aho ibisubizo bihanitse byinjiza birebire byerekana ibimenyetso birebire cyane Amakipe akunze kubona ibisubizo byiza iyo asobanuye ibipimo byujuje ubuziranenge imbere, agakomeza inzira yo kuzamura abantu kubibazo, kandi akurikirana inyungu zibyara umusaruro nibiciro byamakosa mugihe.

Ingaruka & Kurinda

!

Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.

!

Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.

!

Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.

Igishushanyo mbonera

1

Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.

Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

2

Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.

Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

3

Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.

Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

4

Tegura inzira yo gusubiza ibyabaye mbere yo gupima.

Tegura inzira yo gusubiza ibyabaye mbere yo gupima. Fata buri ntambwe nk irembo ryibimenyetso: niba ibipimo bitujujwe, hagarika kuzenguruka, funga icyuho, hanyuma noneho wagure imikoreshereze.

Komeza Ubushakashatsi