Ubuyobozi bwa tekiniki

Urutonde Kuringaniza no Kwitonda Impeta

Urutonde rukurikiranye rugabanya umurongo umwe muremure winjiza murwego rwa GPU nyinshi kurwego rwa token (igihe), kandi Impeta Yitondewe ireka izo GPU zibara kwitondera neza zinyuze urufunguzo / agaciro bahagaritse impeta.

Incamake

Urutonde rukurikiranye rugabanya umurongo umwe muremure winjiza murwego rwa GPU nyinshi kurwego rwa token (igihe), kandi Impeta Yitondewe ireka izo GPU zibara kwitondera neza zinyuze urufunguzo / agaciro bahagaritse impeta. Hamwe na hamwe bakora miriyoni-token imiterere ya windows ishoboka nta GPU nimwe ifata urutonde rwose.

Ikurikiranyabihe hamwe no Kwitondera Impeta ni inyubako ya tekiniki yubaka igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo.

Kwibira cyane

Icyitonderwa gisanzwe gikeneye buri kibazo kugirango ubone urufunguzo / agaciro, bityo ibikorwa byo kwibuka bikura hamwe nuburebure bwikurikiranya kandi K / V yuzuye igomba kuboneka. Ikigereranyo kibangikanye kigabanya urukurikirane kuburyo buri GPU ifite igice gihuza ibimenyetso (nibibazo byabo, urufunguzo, indangagaciro). Impeta yitondere noneho itondekanya GPU mumpeta yumvikana: buri gikoresho gikomeza ibibazo byaho bikosorwa mugihe K / V zahagaritswe hop-by-hop kuzenguruka impeta. Mugihe buri gice kigeze, GPU ibara kwitondera igice kandi ikusanya ibisubizo ukoresheje interineti-softmax (kimwe na max / sum trick nka FlashAttention). Nyuma yumuzingo wuzuye, buri kibazo cyitabiriye urufunguzo rwose, nta GPU yigeze ibika K / V. Byibanze, itumanaho rya K / V ryuzuzanya no kubara, bityo ryongeraho igiciro gito cyurukuta-isaha.

Ubushishozi

Impeta Kwitondera gushingira kuri softmax kumurongo: kwitabwaho birashobora kubarwa guhagarika-guhagarika mugihe ukomeza gukora ntarengwa hamwe nogukora ibisanzwe, hanyuma ugakuraho amafaranga igice cyambere mugihe agaciro kanini kagaragaye. Ibi bituma ibisubizo bibarwa muburyo bwo kwitabwaho byuzuye. Impeta inyura gusa kuri K / V (umunzani ufite ubunini, ntabwo ari urukurikirane rwuzuye), kandi kubera ko itumanaho rya buri hop ryuzuzanya na matmul yabanjirije iyi, umurongo mugari - ntabwo ari kwibuka - bihinduka ibintu bigabanya imipaka.

Kumenya Urutonde Kuringaniza no Kwitonda Impeta

Urutonde rukurikiranye rugabanya umurongo umwe muremure winjiza murwego rwa GPU nyinshi kurwego rwa token (igihe), kandi Impeta Yitondewe ireka izo GPU zibara kwitondera neza zinyuze urufunguzo / agaciro bahagaritse impeta. Hamwe na hamwe bakora miriyoni-token imiterere ya windows ishoboka nta GPU nimwe ifata urutonde rwose. Ikurikiranyabihe hamwe no Kwitondera Impeta ni inyubako ya tekiniki yubaka igira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo. Kugira ngo usobanukirwe byimbitse, fata Urutonde rukurikiranye hamwe nimpeta yitonderwa nkicyitegererezo gikora, ntabwo ari ikintu kimwe: gusobanura ibyagezweho, gusobanura ibitekerezo, no gutandukanya ibyo sisitemu ishobora gukora byizewe nibisaba guca imanza zinzobere.

Mubimenyerezo, amatsinda akomeye akoresha Urutonde rukurikirana hamwe nimpeta Yitondewe yubaka imyubakire, 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 h'uruhererekane Kubangikanya no kwitondera impeta

Iringanizwa ryikurikiranya ririmo kuba ihame ryamahugurwa maremare hamwe nu mwanzuro, akenshi bigahuzwa na tensor hamwe nu muyoboro ugereranije na '4D' cyangwa '5D' ugereranije. Ibihinduka nka striped cyangwa zigzag kwitondera kuringaniza akazi katewe no guhisha impamvu. Tegereza impeta-yamenyeshejwe impeta hejuru ya NVLink no kurushaho guhuza hamwe na KV-cache ikuramo, gusunika uburebure bufatika bugera kuri miriyoni mirongo yikimenyetso cyo kugarura, codebase, hamwe ninyandiko ndende.

Gushyira mu bikorwa Isi

Guhugura 1M-token contexte LLM mugukata buri rutonde kuri 8 GPU hamwe nimpeta

Urutonde rwa Megatron-LM kubangikanya kugabanya ububiko bwibikorwa muri LayeriNorm hamwe nabataye ishuri

Gutunganya igitabo cyose cyangwa ububiko bunini bwa kode muri pass imwe yimbere itagabanije

Guhuza Impeta Kwitonda hamwe na tensor parallelism kugirango ihuze ultra-ndende-imiterere-fatizo yerekana kuri GP-node

Uburyo bwo Gushyira mu bikorwa

Urutonde Kuringaniza no Kwitonda Impeta mubikorwa

Guhugura 1M-token contexte LLM mugukata buri rutonde kuri 8 GPU hamwe nimpeta.

Guhugura imiterere ya 1M-token LLM mugukata buri rutonde kuri 8 GPU hamwe namakipe yitondera Impeta mubusanzwe babona ibisubizo byiza mugihe basobanuye ibipimo byimbere imbere, bagakomeza inzira yo kuzamura abantu kubibazo byimbitse, kandi bagakurikirana inyungu zibyara umusaruro nibiciro byamakosa mugihe.

Urutonde Kuringaniza no Kwitonda Impeta mubikorwa

Urutonde rwa Megatron-LM kubangikanya kugabanya ububiko bwibikorwa muri LayeriNorm hamwe nabataye ishuri.

Urutonde rwa Megatron-LM rugereranya kugabanya ibikorwa byo kwibuka muri LayeriNorm no mu turere twavuye mu matsinda Amakipe akunze kubona umusaruro ushimishije iyo asobanuye ibipimo ngenderwaho biri imbere, agakomeza inzira yo kuzamura abantu kubibazo, kandi akurikirana inyungu zibyara umusaruro hamwe nibiciro byamakosa mugihe runaka.

Urutonde Kuringaniza no Kwitonda Impeta mubikorwa

Gutunganya igitabo cyose cyangwa ububiko bunini bwa kode muri pass imwe yimbere itagabanije.

Gutunganya igitabo cyose cyangwa ububiko bunini bwa kode muri passe imwe yimbere itagabanije Amakipe ubusanzwe 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.

Urutonde Kuringaniza no Kwitonda Impeta mubikorwa

Guhuza Impeta Yitondewe hamwe na tensor parallelism kugirango ihuze ultra-ndende-imiterere-fatizo kuri byinshi-GPU.

Guhuza Impeta Yitondewe hamwe na tensor parallelism kugirango ihuze na ultra-ndende-imiterere-karemano kuri byinshi-GPU node Amakipe ubusanzwe 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.

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