{"id":184029,"date":"2026-10-29T11:40:41","date_gmt":"2026-10-29T09:40:42","guid":{"rendered":"https:\/\/prohoster.info\/blog\/news\/qwenputer-redakcziya-ai-modeli-qwen3-5-sposobnaya-simulirovat-risc-v-cpu-dlya-vypolneniya-koda-na-c-i-micropython"},"modified":"2026-10-29T11:40:46","modified_gmt":"2026-10-29T09:40:46","slug":"qwenputer-a-qwen3-5-ai-model-variant-that-simulates-a-risc-v-cpu-to-run-c-and-micropython-code","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/qwenputer-a-qwen3-5-ai-model-variant-that-simulates-a-risc-v-cpu-to-run-c-and-micropython-code","title":{"rendered":"Qwenputer \u2014 edicioni i modelit AI Qwen3.5, i aft\u00eb t\u00eb simuloj\u00eb CPU RISC-V p\u00ebr t\u00eb ekzekutuar kodin n\u00eb C dhe MicroPython","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
N\u00eb kuad\u00ebr t\u00eb projektit Qwenputer, studiuesit kan\u00eb p\u00ebrgatitur nj\u00eb variant t\u00eb avancuar t\u00eb modelit AI Qwen3.5-4B, n\u00eb t\u00eb cilin \u00ebsht\u00eb implementuar nj\u00eb simulator i plot\u00eb 32-bit CPU bazuar n\u00eb arkitektur\u00ebn RISC-V (RV32IM), q\u00eb mb\u00ebshtet zgjerimin Zba dhe \u00ebsht\u00eb i pajisur me nj\u00eb koprocessor bignum p\u00ebr operacione aritmetike me numra t\u00eb m\u00ebdhenj. Implementimi i b\u00ebrtham\u00ebs CPU \u00ebsht\u00eb kompaktuar n\u00eb nj\u00eb grup peshash dhe \u00ebsht\u00eb integruar n\u00eb nj\u00eb model t\u00eb madh gjuh\u00ebsor n\u00eb form\u00ebn e 11 shtresave shtes\u00eb t\u00eb transformator\u00ebve, t\u00eb shtuar para shtresave kryesore Qwen. Llogaritjet kryhen direkt gjat\u00eb procesit t\u00eb kalimit p\u00ebrpara (forward pass) t\u00eb rrjetit nervor, pa u lidhur me mjete ose plugina t\u00eb jashtme.<\/p>\n
Implementimi kalon me sukses t\u00eb gjitha verifikimet nga grupi zyrtar i testeve riscv-tests dhe \u00ebsht\u00eb n\u00eb gjendje t\u00eb ekzekutoj\u00eb kodin e kompajluar n\u00eb C dhe MicroPython. Kur kryen llogaritje aritmetike me numra t\u00eb m\u00ebdhenj, Qwenputer siguron sakt\u00ebsi prej 97.5-100%, e cila \u00ebsht\u00eb e krahasueshme ose e tejkalon rezultatet e modelit baz\u00eb Qwen3.5 me interpretuesin e jasht\u00ebm Python t\u00eb lidhur. <\/p>\n
<\/center><\/p>\n
P\u00ebr simuluar kujtes\u00ebn operative n\u00eb model, jan\u00eb angazhuar shtresa t\u00eb ve\u00e7anta Gated DeltaNet (RNN lineare), t\u00eb aplikuara n\u00eb modelin standard AI Qwen p\u00ebr ruajtjen e gjendjes s\u00eb ndryshuar gjat\u00eb pun\u00ebs me dritare t\u00eb m\u00ebdha kontekstuale. Duke qen\u00eb se gjendja e der\u00ebs fiksohet dhe aplikohet kodifikimi unitar (one-hot), \u00e7do k\u00ebrkes\u00eb p\u00ebr shkrim riprogramon pik\u00ebrisht nj\u00eb rresht t\u00eb matric\u00ebs s\u00eb gjendjes, duke e b\u00ebr\u00eb t\u00eb mundur q\u00eb gjendja DeltaNet t\u00eb p\u00ebrdoret si memorie e adresueshme me madh\u00ebsi fikse, duke shmangur rritjen lineare t\u00eb cache-it KV (Key-Value) me rritjen e gjat\u00ebsi s\u00eb kontekstit. N\u00eb k\u00ebt\u00eb memorje vendosen regjistrat e CPU, t\u00eb dh\u00ebnat dhe kodi i ekzekutuesh\u00ebm i makineris\u00eb. <\/p>\n
<\/center><\/p>\n
Udh\u00ebzimet e procesorit, si leximi i regjistrave, operacionet aritmetike dhe kalimet kushtore, kompilohet drejtp\u00ebrdrejt n\u00eb peshat e shtresave t\u00eb transformatorit. \u00c7do udh\u00ebzim p\u00ebrputhet me simbolet speciale (ADDR, VAL, STEP), t\u00eb cilat shtresat e modelit AI i interpretojn\u00eb si logjik\u00ebn e p\u00ebrpunimit t\u00eb CPU. P\u00ebr inicializimin e CPU dhe ngarkimin e firmware-it p\u00ebrdoret nj\u00eb drejtues i vog\u00ebl i jasht\u00ebm RvChat, por vet\u00eb kodi i makineris\u00eb p\u00ebrpunon krejt\u00ebsisht brenda peshave t\u00eb modelit.<\/p>\n
N\u00eb kufijt\u00eb e projektit p\u00ebrmendet rreth dyfishi n\u00eb krahasim me modelin origjinal Qwen, me nj\u00eb rritje t\u00eb madh\u00ebsis\u00eb (18.8 GB) duke ruajtur piketat e gjendjes (checkpoint), q\u00eb origjinon kryesisht nga mbushja shtes\u00eb e shtresave t\u00eb reja. Gjithashtu v\u00ebrehet nj\u00eb shpejt\u00ebsi e relativisht ul\u00ebt e llogaritjeve (rreth 3200 tokene n\u00eb sekond\u00eb n\u00eb GPU L4), e cila b\u00ebn q\u00eb llogaritjet e nd\u00ebrlikuara t\u00eb mund t\u00eb zgjasin sekonda, dhe jo milisekonda. <\/p>\n
P\u00ebr m\u00eb tep\u00ebr, \u00ebsht\u00eb realizuar nj\u00eb variant i modelit t\u00eb AI me nj\u00eb sistem eksperimental skedar\u00ebsh, i cili lejon ruajtjen brenda peshave t\u00eb modelit t\u00eb AI deri n\u00eb 512 skeda me 53 mij\u00eb.tokenesh teksti. Nj\u00ebsoj si n\u00eb realizimin e memorjes operative, ruajtja \u00ebsht\u00eb organizuar n\u00eb lidhje me gjendjet DeltaNet, por ndryshon sepse FS ofron ruajtje t\u00eb p\u00ebrhershme, q\u00eb nuk humbet gjendjen midis seancave.<\/p>\n
Qwenputer gjithashtu ka sh\u00ebrbyer si baz\u00eb p\u00ebr studimin e problemeve t\u00eb reja potenciale t\u00eb siguris\u00eb n\u00eb AI. Me shembullin e Qwenputer, \u00ebsht\u00eb treguar se si nj\u00eb model AI i modifikuar mund t\u00eb fsheh\u00eb arsyetimet e saj n\u00eb gjendjen RNN dhe t\u00eb anashkaloj\u00eb mekanizmat e monitorimit t\u00eb arsyetimeve t\u00eb modelit (CoT, Chain-of-Thought). Gjithashtu \u00ebsht\u00eb demonstruar \nmund\u00ebsia e krijimit t\u00eb mekanizmave t\u00eb kontrollit dhe monitorimit t\u00eb siguris\u00eb, t\u00eb integruar n\u00eb peshat e modelit AI, t\u00eb cilat vet\u00eb modeli nuk mund t'i \u00e7aktivizoj\u00eb ose ndryshoj\u00eb. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![](\"\/wp-content\/uploads\/2026\/10\/63b0b8c6c9f2eb8c2e06773ad737106b.png\")
<\/center><\/p>\n![](\"\/wp-content\/uploads\/2026\/10\/5b25214f8fb3789baf16ad3661df0d31.png\")
<\/center><\/p>\n