Overview
Linear tarisiro inotsiva iyo quadratic softmax kutarisisa muTransformers ine math trick inoyera mutsetse nehurefu hwekutevedzana. Performer inzira inofananidzira softmax uchishandisa zvisina kujairika maficha kernels, zvichiita kuti marebesheni akareba akwanise kutenga.
Linear Attention uye Performer Kernels chivakwa chehunyanzvi chinokanganisa mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero.
Deep Dive
Standard Transformer tarisiro inounganidza zvibodzwa pakati pezviviri zvese zvematokeni, inodhura nguva uye ndangariro inokura ine sikweya yekutevedzana kureba (O(n^2)). Linear kutarisisa inonyora zvakare komputa kuti mutengo ukure chete mutsetse (O(n)). Pfungwa yakakosha: softmax attention is softmax(QK^T)V, asi kana ukatsiva softmax nekernel feature mepu phi, unowana phi(Q)(phi(K)^T V). Nekuti kuwanda kwematrix kune mubatanidzwa, unoverengera phi(K)^T V kutanga (diki d-by-d matrix), uchinzvenga hofori n-by-n mamakisi zvachose. Mutambi, anobva ku Google muna 2020, anoita iyi fungidziro yakatendeka yechokwadi softmax uchishandisa FAVOR+ (Fast Attention Via positive Orthogonal Random features), kudhirowa zvisina tsarukano fungidziro inochengeta kernel fungidziro isina rusaruro uye yakagadzikana.
Technical Insight
Performer's FAVOR+ inoyera iyo softmax kernel exp(q.k) ichishandisa zvakanaka zvisina tsarukano maficha: inomepu mibvunzo nemakiyi kuburikidza neyakajairwa Gaussian fungidziro yakaputirwa muexponential, ichivimbisa huremu husiri husina kunaka uye kudzivirira kusamira kwenhamba kwevafungidzi vepakutanga. Kushandisa orthogonal random features kunoderedza kusiyana. Zvine hutsinye, iyo n-by-n yekutarisisa matrix haina kumbobvira yaitwa, saka ndangariro inodonha kubva kune quadratic kuenda kune mutsara, ichigonesa kutevedzana kwemakumi ezviuru zvezviratidzo.
Kubata Linear Kutarisisa uye Muiti Kernels
Linear tarisiro inotsiva iyo quadratic softmax kutarisisa muTransformers ine math trick inoyera mutsetse nehurefu hwekutevedzana. Performer inzira inofananidzira softmax uchishandisa zvisina kujairika maficha kernels, zvichiita kuti marebesheni akareba akwanise kutenga. Linear Attention uye Performer Kernels chivakwa chehunyanzvi chinokanganisa mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero. Kuvaka kunzwisisa kwakadzama, bata Linear Attention uye Performer Kernels semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvaunoda mhedzisiro, kujekesa fungidziro, uye patsanura izvo zvinogona kuitwa nehurongwa nekuvimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Linear Attention uye Performer Kernels inogonesa zvivakwa, data, uye sarudzo dzezvivakwa zvinopesana nekuvimbika uye mutengo. Ivo vanonyora zvakajeka maitiro ebudiriro, bvunzo vachipokana ne data rechokwadi uye mafambiro ebasa, uye iterate zvichibva pane zvakacherechedzwa maitiro ekutadza kwete kuhwina-nguva imwe chete yebhenji. Apa ndipo apo kunzwisisa kwe theoretical kunoshanduka kuve kugona kwakasimba pane chigadzirwa, mutemo, uye mashandiro.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Panguva imwecheteyo, Kukwirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba. Nzira yakatsiga ndeyekubatanidza kukurumidza kuyedza nekutonga: mhanyisa vatyairi vendege, tora humbowo, buritsa matanda esarudzo, uye urambe uchivandudza chengetedzo semaitiro emuenzaniso, zvinotarisirwa nemushandisi, uye zvinodikanwa zvekutonga.
Strategic Impact
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Real-World Implementation
Kugadzira refu genomic kana mapuroteni kutevedzana uko yakazara quadratic kutarisisa kwaizopedza GPU ndangariro
Document-level muchidimbu pamusoro pemishumo yakareba kwazvo pasina chunking, uchishandisa Performer-style musana
Inoshanda-refu-fomu redhiyo kana nguva-yakatevedzana modhi uko kutevedzana kunotora makumi ezviuru zvenhanho
Kuderedza mutengo wekufungidzira mune refu-chinyorwa chat modhi nekutsiva mamwe softmax layer ane mutsara-yekutarisa akasiyana.
Maitiro Ekuita
Linear Attention uye Performer Kernels mukuita
Kugadzira refu genomic kana mapuroteni kutevedzana uko yakazara quadratic kutarisisa kwaizopedza GPU ndangariro.
Kugadzira marefu genomic kana mapuroteni kutevedzana uko kuzere kwequadratic kutarisisa kwaizopedza GPU ndangariro Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Linear Attention uye Performer Kernels mukuita
Document-level muchidimbu pamusoro pemishumo yakareba kwazvo pasina chunking, uchishandisa Performer-style musana.
Document-level pfupiso pamusoro pemishumo yakareba kwazvo pasina chunking, uchishandisa Performer-style musana Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Linear Attention uye Performer Kernels mukuita
Inoshanda-refu-fomu redhiyo kana nguva-yakatevedzana modhi uko kutevedzana kunotora makumi ezviuru zvenhanho.
Inoshanda-yakareba-fomu redhiyo kana nguva-yakatevedzana modhi uko kutevedzana kunosvika makumi ezviuru zvenhanho Zvikwata zvinowanzowana mhedzisiro iri nani kana vachitsanangudza zvemhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye kukanganisa mutengo nekufamba kwenguva.
Linear Attention uye Performer Kernels mukuita
Kuderedza mutengo wekufungidzira mumamodhi akareba-echinyorwa chekutaura nekutsiva mamwe akapfava max nemutsara-yekutarisa akasiyana.
Kuderedza mutengo wekufungidzira mumhando dzenguva refu dzekutaura nekutsiva mamwe maSoftmax masiketi ane mutsara-yekutarisisa akasiyana Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Njodzi & Guardrails
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.
Implementation Roadmap
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Benchmark pasi pechokwadi mutoro uye data mamiriro.
Benchmark pasi pechokwadi mutoro uye data mamiriro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.