Overview
Kudhirodha kwekufungidzira kunoshandisa diki, inokurumidza 'dhizaini' modhi kufungidzira akati wandei ari kuuya ma tokeni ayo hombe modhi inozo ongorora mupasi rimwe. Inomhanyisa chizvarwa chemavara 2-3x pasina shanduko kune inobuda.
Inofungidzira Decoding Draft Models chikamu chemutauro-AI stack inoshandiswa kuverenga, kugadzira, kuronga, uye kushandura zvinyorwa uye kutaura pamwero.
Deep Dive
Mienzaniso mikuru yemitauro inoburitsa chinyorwa chimwe chete panguva, uye nhanho imwe neimwe inoda kuenderera mberi kwakazara nemabhiriyoni emaparamita - inononoka uye inosungwa mundangariro. Yekufungidzira decoding inorwisa izvi nekubatanidza iyo hombe 'chinangwa' modhi neyakachipa 'dhizaini' modhi. Iyo yekudhirowa modhi inokurumidza kukurudzira chunk ye, toti, 4-8 mumiriri tokens. Iyo hombe modhi inozoagadzirisa ese mune imwechete yakafanana kuenda mberi uye inotarisa imwe neimwe. Zviratidzo zvinopindirana neizvo mufananidzo mukuru ungadai wakagadzira zvinogamuchirwa; yekutanga kusawirirana inogadziriswa uye imwe yese inoraswa. Nekuti kuonesa tokeni akati wandei kamwechete kunodhura zvakangofanana nekugadzira imwe, inotambirwa inomhanya inenge yemahara. Zvine hutsinye, nhanho yekuramba-sampling inovimbisa kugovera kwekupedzisira kwakafanana nekumhanyisa iyo hombe modhi chete - kumhanya pasina kurasikirwa kwemhando.
Technical Insight
Iyo yakakosha trick ndeye yakagadziridzwa yekuramba-sampling bvunzo. Pachiratidzo chega chega chakanyorwa, mukana weiyo inotarirwa modhi inofananidzwa neyeiyo dhizaini modhi. Kana chinangwa chinopa mukana wakaenzana kana wepamusoro, chiratidzo chinogamuchirwa; zvikasadaro inogamuchirwa ine mukana wakaenzana nereshiyo, uye pakurambwa chiratidzo chakagadziriswa chinotorwa kubva kune yakagadziridzwa yasara kugovera. Masvomhu aya anoita kuti zvinobuda zvienzane nesampling zvakananga kubva kune hombe modhi.
Mastering Yekufungidzira Decoding Draft Models
Kudhirodha kwekufungidzira kunoshandisa diki, inokurumidza 'dhizaini' modhi kufungidzira akati wandei ari kuuya ma tokeni ayo hombe modhi inozo ongorora mupasi rimwe. Inomhanyisa chizvarwa chemavara 2-3x pasina shanduko kune inobuda. Inofungidzira Decoding Draft Models chikamu chemutauro-AI stack inoshandiswa kuverenga, kugadzira, kuronga, uye kushandura zvinyorwa uye kutaura pamwero. Kuti uvake kunzwisisa kwakadzama, tora maSpeculative Decoding Draft Models semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvinodiwa, kujekesa fungidziro, uye patsanura izvo zvinogona kuitwa nehurongwa hwakavimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Speculative Decoding Draft Models dhizaini yekukurudzira, kudzoreredza, uye kuongorora zvishwe seimwe yakabatanidzwa yekutaurirana system. 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.
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana. Panguva imwecheteyo, chokwadi cheHallucified chinogona kupinda chinyararire mishumo, kuyerera kwetsigiro, kana kutsvagisa zvinobuda. 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
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana.
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Inopamhidzira kupinda mumitauro yese nemataera ekutaurirana.
Inopamhidzira kupinda mumitauro yese nemataera ekutaurirana. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Zvikwata zvinogona kupedza nguva yakawanda pakutonga uku otomatiki ichibata kudzokorora.
Zvikwata zvinogona kupedza nguva yakawanda pakutonga uku otomatiki ichibata kudzokorora. 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
Anthropic, OpenAI, uye Google vanoshandisa decoding yekufungidzira kuderedza kunonoka uye kushandisa mutengo kune vabatsiri vekutaura vanoshandira mamirioni evashandisi.
vLLM uye NVIDIA TensorRT-LLM ngarava yakavakirwa-mukati yekufungidzira decoding kuitira kuti ivo vanozvitambira vagone kukurumidzira Llama kana Mistral kutumirwa.
Kubatanidza 7B dhizaini modhi ine 70B yakanangana (semuenzaniso, Llama-3 mhuri) kune inokwana kaviri tokens-pasekondi pane imwe GPU.
Maturusi ekupedzisa kodhi anoshandisa diki dhizaini modhi kuratidza boilerplate iyo yakakura modhi inosimbisa, ichichengeta mazano ari nyore mupepeti.
Maitiro Ekuita
Inofungidzira Decoding Draft Models mukuita
Anthropic, OpenAI, uye Google vanoshandisa decoding yekufungidzira kuderedza kunonoka uye kushandisa mutengo kune vabatsiri vekutaura vanoshandira mamirioni evashandisi.
Anthropic, OpenAI, uye Google shandisa fungidziro decoding kuderedza latency uye kuendesa mutengo kune vabatsiri vechat vanoshandira mamirioni evashandisi Zvikwata zvinowanzowana mibairo iri nani kumberi, chengetedza nzira yekukwira kwechigadzirwa uye kuronda kukwira kwechigadzirwa.
Inofungidzira Decoding Draft Models mukuita
vLLM uye NVIDIA TensorRT-LLM ngarava yakavakirwa-mukati yekufungidzira decoding kuitira kuti ivo vanozvitambira vagone kukurumidzira Llama kana Mistral kutumirwa.
vLLM neNVIDIA TensorRT-LLM ngarava yakavakirwa-mukati yekufungidzira decoding kuitira kuti ivo vanozvitambira vagone kukurumidzira Llama kana Mistral deployments 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.
Inofungidzira Decoding Draft Models mukuita
Kubatanidza 7B dhizaini modhi ine 70B yakanangana (semuenzaniso, Llama-3 mhuri) kune inokwana kaviri tokens-pasekondi pane imwe GPU.
Kubatanidza 7B dhizaini modhi ine 70B yakanangana (semuenzaniso, mhuri yeLlama-3) kuti ingangoita kaviri tokens-per-sekondi pane imwechete GPU 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.
Inofungidzira Decoding Draft Models mukuita
Maturusi ekupedzisa kodhi anoshandisa diki dhizaini modhi kuratidza boilerplate iyo yakakura modhi inosimbisa, ichichengeta mazano ari nyore mupepeti.
Maturusi ekupedzisa kodhi anoshandisa diki dhizaini modhi kuratidza boilerplate iyo yakakura modhi inosimbisa, kuchengetedza mazano ari nyore mupepeti Matimu anowanzo kuwana mibairo iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye kukanganisa mutengo nekufamba kwenguva.
Njodzi & Guardrails
Chokwadi chehuroyi chinogona kupinda chinyararire mishumo, kuyerera kwetsigiro, kana tsvakiridzo.
Kunzwa nekukasira kunogona kugadzira mhedzisiro isingaenderane pane zvikumbiro zvakafanana.
Sensitive text data inogona kuburitswa kana zvidhiraivho zvisina kusimba.
Implementation Roadmap
Tsanangura chimiro chekubuda, toni, uye mhando zviyero usati waburitsa.
Tsanangura chimiro chekubuda, toni, uye mhando zviyero usati waburitsa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Mhinduro dzepasi neakavimbika masosi pese pazvine basa.
Mhinduro dzepasi neakavimbika masosi pese pazvine basa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Chengetedza ongororo yekuongorora yemunhu kune yakakwira-stake zvinobuda.
Chengetedza ongororo yekuongorora yemunhu kune yakakwira-stake zvinobuda. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Tevera maitiro ekutadza uye dzidzisazve kukurudzira kana mafambiro ebasa nguva nenguva.
Tevera maitiro ekutadza uye dzidzisazve kukurudzira kana mafambiro ebasa nguva nenguva. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.