Ubuyobozi bwa tekiniki

Byuzuye Byuzuye Ibyatanzwe

Byuzuye Byuzuye Data Parallel (FSDP) nubuhanga bukwirakwizwa bwamahugurwa agabanya ibipimo byikitegererezo, gradients, hamwe na optimizer leta muri GPU nyinshi kuburyo buri gikoresho gifata igice gusa.

Incamake

Byuzuye Byuzuye Data Parallel (FSDP) nubuhanga bukwirakwizwa bwamahugurwa agabanya ibipimo byikitegererezo, gradients, hamwe na optimizer leta muri GPU nyinshi kuburyo buri gikoresho gifata igice gusa. Bituma imyitozo nini nini ishoboka kubikoresho bidashobora na rimwe guhuza moderi yose mububiko bumwe bwa GPU.

Byuzuye Byuzuye Data Parallel nububiko bwa tekiniki yububiko bugira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo.

Kwibira cyane

Imibare gakondo ibangikanye ibika kopi yuzuye yicyitegererezo kuri buri GPU, isesagura ububiko na caps ingano yubunini. FSDP, ikunzwe na __AIU_PROTECTED_13 __ 'PyTorch kandi ihumekwa na ZeRO ya __AIU_PROTECTED_5 __, ahubwo igabanya ibintu bitatu mubikoresho: ibipimo, gradients, hamwe na leta nziza. Mugihe cyo gutambuka imbere, buri GPU ikusanya by'agateganyo uburemere bwuzuye kurwego irimo kubara ikoresheje byose-ikora, ikora ibarwa, hanyuma ihita irekura kopi yakusanyije. Inyuma yinyuma ikora kimwe, ikurikirwa no kugabanya-gutatanya gukwirakwiza uduce duto duto dusubira kuri GPU zabo. Kuberako buri gikoresho kibika burundu igice cyicyitegererezo, gukoresha kwibuka bitonyanga hafi yumubare wa GPUs, ukareka amakipe akamenyereza moderi hamwe na miliyari icumi cyangwa magana.

Ubushishozi

FSDP icuruza itumanaho ryiyongera kubitsa. Ibipimo bya buri cyiciro byongeye kubakwa kubisabwa hamwe byose hamwe mbere yo gukoreshwa hanyuma bikajugunywa nyuma, mugihe gradients zahujwe kandi zigabanijwe kugabanya-gutatanya. Itumanaho rirashobora guhuzagurika no kubara muguhitamo ibipimo bikurikiraho mugihe urwego rwubu rukora, guhisha byinshi mubukererwe bwurusobe. Kuringaniza granularity (politike yo gupfunyika) iringaniza ikirenge cyo kwibuka hamwe n'itumanaho hejuru.

Kumenya neza amakuru asangiwe neza

Byuzuye Byuzuye Data Parallel (FSDP) nubuhanga bukwirakwizwa bwamahugurwa agabanya ibipimo byikitegererezo, gradients, hamwe na optimizer leta muri GPU nyinshi kuburyo buri gikoresho gifata igice gusa. Bituma imyitozo nini nini ishoboka kubikoresho bidashobora na rimwe guhuza moderi yose mububiko bumwe bwa GPU. Byuzuye Byuzuye Data Parallel nububiko bwa tekiniki yububiko bugira ingaruka kumiterere yicyitegererezo, igiciro cyibikorwa remezo, ubukererwe, no kwizerwa mubipimo. Kugirango wubake byimbitse, fata amakuru yuzuye asanganywe nkurugero rwimikorere, ntabwo arikintu kimwe: gusobanura ibyagezweho, gusobanura ibitekerezo, no gutandukanya ibyo sisitemu ishobora gukora byizewe nibisaba guca imanza zinzobere.

Mubimenyerezo, amatsinda akomeye akoresha amakuru yuzuye asanganywe Kuringaniza yubaka ubwubatsi, amakuru, 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 huzuye amakuru asanganywe

FSDP ihinduka nkibisanzwe byugururiwe amahugurwa manini-manini, hamwe na FSDP2 muri PyTorch itezimbere imikoreshereze hamwe na buri kintu cyerekana. Witegereze kurushaho guhuza hamwe na tensor hamwe nuyoboro ugereranije kubintu bya tiriyari-ibipimo byerekana, inkunga nziza yo kuvanga neza na fp8, hamwe no gufunga ubwenge byikora bitoranya imipaka kuri wewe. Mugihe imiyoboro ya GPU ihuza nka NVLink na InfiniBand igenda yihuta, igiciro cyitumanaho rya sharding gikomeza kugabanuka, bigatuma gifatika mubipimo binini cyane.

Gushyira mu bikorwa Isi

Kuringaniza neza miliyari 70-yerekana ibipimo bya Llama muri 8 GPUs kugiti cye kidashobora gufata uburemere bwuzuye.

Gutegura imvugo nini yindimi muri laboratoire ya AI mugukata ibintu byiza (biganje mububiko hamwe na Adam) mumajana yihuta.

Abashakashatsi bifashishije igikoresho cya FSDP cya PyTorch kugirango bahugure abahindura icyerekezo kuri cluster ya kaminuza bataguze 80GB GPUs.

Gukomatanya FSDP hamwe na mix-precision bfloat16 kugirango ugabanye hafi kwibuka no kwihutisha amahugurwa kuri moderi nyinshi.

Uburyo bwo Gushyira mu bikorwa

Byuzuye Byagabanijwe Byuzuye Kuringaniza mubikorwa

Kuringaniza neza miliyari 70-yerekana ibipimo bya Llama muri 8 GPUs kugiti cye kidashobora gufata uburemere bwuzuye.

Kuringaniza neza miliyari 70-yerekana urugero rwa Llama muri 8 GPUs ku giti cye idashobora kwihanganira uburemere bwuzuye 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.

Byuzuye Byagabanijwe Byuzuye Kuringaniza mubikorwa

Gutegura imvugo nini yindimi muri laboratoire ya AI mugukata ibintu byiza (biganje mububiko hamwe na Adam) mumajana yihuta.

Gutegura imvugo nini yindimi muri laboratoire ya AI ukoresheje uburyo bwiza (bwiganjemo kwibuka hamwe na Adam) mumagana yihuta yamakipe Amakipe akunze kubona ibisubizo byiza iyo asobanuye ibipimo byimbere imbere, agakomeza inzira yo kuzamura abantu kubibazo byimbere, kandi akurikirana inyungu zibyara umusaruro nibiciro byigihe.

Byuzuye Byagabanijwe Byuzuye Kuringaniza mubikorwa

Abashakashatsi bifashishije igikoresho cya FSDP cya PyTorch kugirango bahugure abahindura icyerekezo kuri cluster ya kaminuza bataguze 80GB GPUs.

Abashakashatsi bifashishije impuzu ya FSDP ya PyTorch kugirango bahugure abahindura iyerekwa kuri cluster ya kaminuza bataguze ibendera rya 80GB GPUs Amakipe akunze kubona umusaruro ushimishije mugihe basobanuye ibipimo byujuje ubuziranenge imbere, bagakomeza inzira yo kuzamuka kwabantu kubibazo bikabije, kandi bagakurikirana inyungu zibyara umusaruro nibiciro byigihe.

Byuzuye Byagabanijwe Byuzuye Kuringaniza mubikorwa

Gukomatanya FSDP hamwe na mix-precision bfloat16 kugirango ugabanye hafi kwibuka no kwihutisha amahugurwa kuri moderi nyinshi.

Guhuza FSDP hamwe na bfloat 16 ivanze kugirango igabanye hafi kwibuka no kwihutisha kwinjiza muma moderi ya multimodal Amakipe ubusanzwe abona ibisubizo byiza iyo asobanuye ibipimo byimbere imbere, agakomeza inzira yo kuzamura abantu kubibazo byimbere, kandi akurikirana inyungu zibyara umusaruro nibiciro byigihe.

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