Moteri ya MDST: Koresha Moderi ya GGUF muri Browser hamwe na WebGPU / WASM

Moteri ya MDST nigihe kigenda gifasha abitezimbere nubucuruzi gukora imiterere nini ya GGUF yerekana imiterere nini yindimi imbere muri mushakisha ukoresheje WebGPU na WebAssemble (WASM), bikuraho ibikenewe kuri seriveri yabugenewe cyangwa igicu GPU. Ihinduka ryerekeranye nabakiriya-bonyine AI yerekana imyandikire yerekana uburyo ibintu byubwenge bitangwa murubuga rwa interineti, bigatuma AI yihariye, itinda-bwihuse AI igera kubantu bose bafite mushakisha igezweho.

Niki Cyane Moteri ya MDST kandi Kuki Bitwaye?

Moteri ya MDST ni mushakisha-kavukire ya AI yerekana uburyo bwo gupakira no gukoresha moderi ya GGUF igereranijwe - imiterere imwe ikunzwe n'imishinga nka llama.cpp - muburyo butaziguye kurubuga. Aho kugirango uhindure icyifuzo cyose cya AI ukoresheje igicu cyanyuma, MDST ikora umwanzuro wicyitegererezo kubikoresho byumukoresha wenyine ukoresheje WebGPU API ya mushakisha ya GPU yihuta yo kubara hamwe na WebAssemble kugirango ibikorwa bya CPU bigaruke.

Ibi bifite akamaro kanini kubwimpamvu nyinshi. Ubwa mbere, ikuraho uruzinduko-rugendo rwihuta rwerekeranye na seriveri kuruhande. Icya kabiri, ibika amakuru yumukoresha yuzuye kubikoresho, ninyungu yibanga ryibikorwa bya entreprise hamwe nabaguzi kimwe. Icya gatatu, iragabanya cyane ibikorwa remezo kubucuruzi bwakwishyura ubundi guhamagara API cyangwa kugumana amahuriro yabo ya GPU.

"

Nigute WebGPU na WASM Bituma Muri-Browser AI bishoboka?

Gusobanukirwa ishingiro rya tekiniki ya MDST Moteri bisaba kureba muri make ibice bibiri byibanze bya mushakisha ikoresha. WebGPU niyo isimbuye WebGL, itanga urwego rwo hasi rwa GPU ruva muri JavaScript na kode ya shader ya WGSL. Bitandukanye nabayibanjirije, WebGPU ishyigikira igicucu cya compte, arizo mikorere yimirimo yo kugwiza matrix yiganjemo LLM. Ibi bivuze ko MDST ishobora kohereza ibikorwa bya tensor muri GPU muburyo bubangikanye cyane, kugera kubintu byinjira mbere bitashobokaga imbere muri sandbox ya mushakisha.

WebAssemble ikora nkibisubira inyuma hamwe nintego yo gukusanya intego ya moteri yimikorere ya logique. Kubikoresho bidafite inkunga ya WebGPU - mushakisha zishaje, ibidukikije bimwe na bimwe bigendanwa, cyangwa ibizamini byo kutagira umutwe - WASM itanga imikorere ikora, igendanwa ikoreshwa na C ++ cyangwa Rust code yakozwe ku muvuduko urenze kure JavaScript isanzwe. Hamwe na hamwe, WebGPU na WASM bagize ingamba zo gushyira mu bikorwa: GPU-mbere iyo iboneka, CPU-ikoresheje-WASM iyo itari.

Moderi ya GGUF niyihe kandi ni ukubera iki iyo format iri hagati yubu buryo?

GGUF (Imiterere ihuriweho na GPT) nuburyo bubiri bwa dosiye igizwe nuburemere bwikitegererezo, amakuru ya tokenizer, na metadata mubikoresho bimwe byoroshye. Ubusanzwe byateguwe kugirango bishyigikire neza muri llama.cpp, GGUF yabaye de-facto yuburyo bwa moderi yuzuye ifunguye-uburemere kuko ishyigikira urwego rwinshi rwo kubara-kuva kuri 2-bit kugeza kuri 8-bit-kwemerera abitezimbere guhitamo ubucuruzi hagati yubunini bwikitegererezo, ikirenge cyibikoresho, nibisohoka.

Kubisobanuro bishingiye kuri mushakisha, kubara ntabwo ari ubushake-ni ngombwa. Moderi yuzuye ya 7B isaba hafi 14 GB yo kwibuka. Ku mubare wa Q4, iyo moderi imwe igabanuka kugera kuri GB 4, naho kuri Q2 irashobora kugabanuka munsi ya 2 GB. Inkunga ya moteri ya MDST kuri GGUF bivuze ko abitezimbere bashobora gukoresha mu buryo butaziguye urusobe runini rw’ibinyabuzima bimaze kugereranywa nta ntambwe yongeyeho yo guhindura, bikagabanya cyane inzitizi yo kwishyira hamwe.

Nibihe Byukuri-Gukoresha Imanza Kubucuruzi Bwerekana Moderi ya GGUF muri Browser?

Porogaramu ifatika ya-mushakisha GGUF yerekana hafi yinganda zose. Abashoramari bakoresha ubu buryo bwo gufungura ubushobozi bwahoze bubuza ibiciro cyangwa ubuzima bwite-budahuye nibisubizo bya AI. Imanza zingenzi zikoreshwa zirimo:

• Abafasha ba AI bashoboye:
• Isesengura ryinyandiko yihariye: Amategeko yemewe, ubuvuzi, nubukungu byakazi aho inyandiko zoroshye zitagomba kuva mubikoresho byumukoresha, nyamara zikungukirwa nincamake zikoreshwa na AI.