Modèle mixtral ak sparse
Mixtral mooy xeetu njaxasu eksper yu Mistral AI biy joxe kalite model bu mag ci gaawaay model bu ndaw.
Résumé
Sparse models like it activate only a fraction of their parameters per token, cutting compute without sacrificing capability.
Plongeur bu xóot
Mixtral 8x7B, bi Mistral AI genne ci 2023, siiwal njaxasu eksper yi (MoE) ci xeetu ubbeeku yi. Dafa am juróom ñatti 'expert' reso feed-forward yu wuute ci layer bu nekk, ak luy tollu ci 47 milyaar ciy parametre, waaye routeur bu woyof bi dafay tànnee ñaari expert ci token bu nekk. Ci noonu la tollu ci 13 milyaari parametre kese lañuy dox ci token bu nekk, kon inference dafay daw lu gaaw ni model dense 13B ci di yegg ci kalite bu méngoo ak yeneen yu gëna mag. Mixtral méngoo wala raw GPT-3.5 ak Llama 2 70B ci benchmark yu bari, fekk moo gëna gaaw te gëna yomb ci liggéey. Ginaaw loolu Mistral génne Mixtral 8x22B. Modèle bi dafa am lisaas bu ubbeeku ci Apache 2.0, di gaawa jëfandikoo ak defar bu baax ci askanu open-source bi.
Gis-gis xarala
Ci biir MoE bu bariwul, dañuy wecci blok feed-forward bu dëgër bi ak N reso expert boole ci reso gating bu ndaw (routeur bi). Ci token bu nekk, routeur bi dafay xayma poñ yi ba noppi tànn top-k expert yi (top-2 ci Mixtral), di router token bi ci ñoom kese. Seen génnekaay dañu leen di pondére ba noppi boole leen. Ndax li gëna bari ci kàngam yi duñu def dara ci token bu nekk, model bi dafay tëye ay parametre yu bari ci mémoire bi waaye di def calcul yu néew. Kompromis bi: bépp kàngam dafa wara nekk ci VRAM doonte yenn ci ñoom kese lañuy daw.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu xeetu Mixtral ak Sparse
Sparse MoE mingi leegi ci diggu IA. Xaarandi yeneen MoE yu ubbeeku, yoon yu gëna sew ak eksper yu ndaw yu bari, ak jëmmal eksper yu bokk wala hybrid yuy gëna yokk njariñ li. Ginaaw model yi dañuy dem ba ci ay trillion ci parametre yi, sparsity mooy levier bi gëna am solo ngir tëye inference bi ci anam wu woyof. Gëstu mingi wax ci barab yu néew doole yi ci MoE, balance charge ci diggante eksper yi, memory bi ci kaw, ak stabilite tàggat yaram, ci noonu la hardware ak stack yiy liggéey di gëna mëna optimiser gëna jëm ci eksper yi.
Doxal ci àdduna dëgg
Liggéeyukaay chatbot bu baax ci njëg ak gaawaay bu gëna ndaw
Hosting sa bopp ab xeetu Apache-2.0 bu am lisaas ngir ay produit njaay te doo fay jëfandikoo
Defar bu baax jeffin ju benn nit ci Mixtral ngir kodage, tënk, wala liggéey ci làkk yu bari
Dawal inference bu gaaw ci benn serwër bu bari GPU fu model bu dëgër 70B di yeex lool
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Mixtral and Sparse Models quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Model ak paralelismu tuyo
Laaj yi ñuy faral di laaj
What is Mixtral and Sparse Models?
Mixtral mooy xeetu njaxasu eksper yu Mistral AI biy joxe kalite model bu mag ci gaawaay model bu ndaw. Modèle yu néew yu mel ni ñoom duñu aktive benn wàll ci seeni parametre ci token bu nekk, dagg ordinatër te duñu yàq kàttan gi.
Ci Mixtral 8x7B, ñaata kàngam ñooy liggéey ci token bu nekk?
Mixtral dafay jëfandikoo top-2 routing: router bi dafay tànn ñaar ci juróom ñatti kàngam yi ngir liggéey ci token bu nekk.
Ban cër mooy dogal ban kàngam lañuy yónnee token bi?
Benn reso gating bu ndaw, ñu koy woowe routeur, mooy joxe poñ ci kàngam yi ba noppi tànn kiy jëfandikoo token bu nekk.
Doonte Mixtral 8x7B amna lu tollu ci 47B ci parametre yi, lu tollu ci ñaata lañu ci token bu nekk?
Ndax ñaari kàngam kese ñooy daw ci token bu nekk, lu tollu ci 13 milyaari parametre ñooy dox, loolu moo koy jox gaawaayu model bu gëna ndaw.
Lan mooy njaay bu am solo ci xeetu MoE yu bari yu melni Mixtral?
MoE dafay denc calcul ci token bu nekk waaye mingi sàkku ci kàngam yépp ñu tëye leen ci VRAM, di yokk soxlay mémoire.
Ci ban lisence la Mistral AI génne Mixtral, tax ñu nangu ko?
Mixtral ñu ngi ko génne ci ndigalu Apache 2.0, muy may ñu jëfandikoo ko ci jënd ak jaay te doo fay.