InoteveraGaidhi rinotevera
Tensor Parallelism yeMakuru Models
Tekinoroji
Nhungamiro yehunyanzvi
Model uye Pipeline Parallelism
Kana modhi yakakurisa kuti ikwane pane imwe GPU, modhi uye pombi parallelism inopatsanura modhi yacho pachayo pamidziyo.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
Pa peji ino2 min verenga
Pfupiso
Izvi ndizvo zvinoita kuti kudzidzisa mamodheru emitauro mikuru ane mazana emabhiriyoni emaparamita anogona kuitika.
Kudzika Kwakadzika
Model parallelism partitions modhi imwechete kune akawanda maGPU saka hapana mudziyo mumwe unoda kubata huremu hwese. Pane zvinonaka zviviri zvikuru. Tensor (intra-layer) parallelism inotsemura math mukati mechikamu, sekucheka matrix akakura kuwanda muGPUs iyo yega yega inoverengera chikamu chekubuda. Pipeline (inter-layer) parallelism inopa akasiyana akateedzana akateedzana kune akasiyana maGPU, saka layer block 1 inogara paGPU 0, block 2 paGPU 1, zvichingodaro, ne activation inopfuudzwa kumberi semutsetse wegungano. Dambudziko nepipelini yekusaziva ndiyo 'bubble': nepo GPU 0 inoshanda pabatch yekutanga, yakadzika maGPU anogara asina basa. Pipelining inotsemura batch yega yega kuita madiki-mabhechi kuti ese matanho agare akabatikana, zvinoshamisa kuvandudza mashandisirwo.
Kuitwa muNyika yeChokwadi
Kudzidzira GPT-maitiro emhando neNVIDIA Megatron-LM, iyo inotsemura kutarisa kwega kwega kweshanduko uye kudyisa-mberi matrices paGPUs kuburikidza ne tensor parallelism.
Kushandisa GPipe kuisa zvikamu zvakasiyana zvechiono chihombe kana modhi yemutauro pane akasiyana accelerator nepo micro-batching ichivachengeta vakabatikana.
DeepSpeed's pombi injini yekuparadzanisa mazana-mazana-mabhiriyoni-parameter modhi mumatanho munzvimbo dzakawanda.
Kubatanidza tensor parallelism mukati meimwe 8-GPU sevha ine pombi parallelism inotenderera akawanda maseva kudzidzisa modhi yakakura kwazvo pamushini mumwe.
Ramangwana reMuenzaniso uye Pipeline Parallelism
Mafuremu anowedzera otomatiki dambudziko rakaoma rekusarudza nzira yekugovanisa modhi pamidziyo yese, uchishandisa profiling uye kutsvaga kuenzanisa komputa nekutaurirana. Tarisira kusanganisa kwakasimba kwe tensor, pombi, uye parallelism yedata (3D parallelism), yakangwara micro-batch kuronga kupotsa kubvisa pombi mabhubhu, uye Hardware ine kukurumidza kubatanidza saka kupatsanura imwe layer kune machipisi kunove kuchipa uye kuwanda maitiro kune anogara akakura modhi.
Kuzvishandisa
Nei zvichikosha
- Mutengo uye bhajeti. Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
- Sarudzo dzakajeka. Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
- Kudzora kwemhando yepamusoro. Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Chenjera
- Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
- Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
- Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.
Matanho anotevera
- Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
- Benchmark pasi pechokwadi mutoro uye data mamiriro.
- Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
- Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Ramba Uchiongorora
Tsamba yemahara
Tora iyo yemazuva ese AI pfupiso
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Take the Model and Pipeline Parallelism quiz
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Mibvunzo inowanzo bvunzwa
Chii chinonzi Model uye Pipeline Parallelism?
Kana modhi yakakurisa kuti ikwane pane imwe GPU, modhi uye pombi parallelism inopatsanura modhi yacho pachayo pamidziyo. Izvi ndizvo zvinoita kuti kudzidzisa mamodheru emitauro mikuru ane mazana emabhiriyoni emaparamita anogona kuitika.
Nderipi dambudziko rakakosha rinogadziriswa modhi uye pombi parallelism?
Modhi uye pombi parallelism kupatsanura modhi pachayo pamidziyo, ichigonesa kudzidziswa kwemanetiweki akakura kwazvo kupfuura chero GPU imwe chete yaigona kubata.
Ko tensor (intra-layer) parallelism inopatsanura sei?
Tensor parallelism inokamura computation mukati mechikamu chimwe chete, semuenzaniso kupatsanura huremu hombe matrix kuitira kuti GPU yega yega iverenge chikamu chekubuda.
Chii chinonzi 'bubble' mune naive pipeline parallelism?
Pakutanga kwepaipi nhanho, nhanho dzepasi pemvura hadzina chekupinza uye gara zvisina basa, kutambisa GPU nguva; gap iri risina basa rinonzi bubble.
Kufanana kwepombi kunoderedza sei bubble?
Kugovanisa batch kuita madiki-mabhechi kunoita kuti akawanda madiki-mabhechi atore matanho akasiyana panguva imwe chete, achichengeta ese maGPU akabatikana uye kupera nguva yekusaita basa.
Sei tensor parallelism ichiwanzo chengetwa mukati meimwe node?
Tensor parallelism inotaurirana kazhinji kusanganisa mhedzisiro yematrix, saka inoiswa painobatana bandwidth yakakwirira, mukati menode pamusoro peNVLink.
Ramba uchidzidza
Mazano anoenderana
Mamwe madhairekitori akasarudzirwa nyaya iyi