{"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\/cs\/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 edice AI modelu Qwen3.5, schopn\u00e1 simulovat CPU RISC-V pro vykon\u00e1v\u00e1n\u00ed k\u00f3du v C a MicroPythonu","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
V r\u00e1mci projektu Qwenputer v\u00fdzkumn\u00edci p\u0159ipravili roz\u0161\u00ed\u0159enou variantu AI modelu Qwen3.5-4B, ve kter\u00e9 je implementov\u00e1n vestav\u011bn\u00fd simul\u00e1tor plnohodnotn\u00e9ho 32bitov\u00e9ho CPU na b\u00e1zi architektury RISC-V (RV32IM), podporuj\u00edc\u00edho roz\u0161\u00ed\u0159en\u00ed Zba a vybaven\u00e9ho koprocesorem bignum pro aritmetick\u00e9 operace s velk\u00fdmi \u010d\u00edsly. Implementace j\u00e1dra CPU byla zkompilov\u00e1na do sady vah a integrov\u00e1na do velk\u00e9ho jazykov\u00e9ho modelu jako 11 dodate\u010dn\u00fdch vrstev transform\u00e1tor\u016f, kter\u00e9 byly p\u0159id\u00e1ny p\u0159ed hlavn\u00edmi vrstvami Qwen. V\u00fdpo\u010dty se prov\u00e1d\u011bj\u00ed p\u0159\u00edmo v procesu p\u0159\u00edm\u00e9ho pr\u016fchodu (forward pass) neuronov\u00e9 s\u00edt\u011b, bez p\u0159ipojen\u00ed extern\u00edch n\u00e1stroj\u016f nebo plugin\u016f.<\/p>\n
Implementace \u00fasp\u011b\u0161n\u011b pro\u0161la v\u0161emi testy z ofici\u00e1ln\u00ed sady test\u016f riscv-tests a je schopn\u00e1 vykon\u00e1vat zkompilovan\u00fd k\u00f3d na C a MicroPython. P\u0159i prov\u00e1d\u011bn\u00ed aritmetick\u00fdch v\u00fdpo\u010dt\u016f s velk\u00fdmi \u010d\u00edsly Qwenputer zaji\u0161\u0165uje p\u0159esnost 97,5\u2013100 %, co\u017e je srovnateln\u00e9 nebo p\u0159ekra\u010duje v\u00fdsledky z\u00e1kladn\u00edho modelu Qwen3.5 s p\u0159ipojen\u00fdm extern\u00edm interpretem Python. <\/p>\n
<\/center><\/p>\n
Pro simulaci opera\u010dn\u00ed pam\u011bti v modelu jsou vyu\u017e\u00edv\u00e1ny samostatn\u00e9 vrstvy Gated DeltaNet (line\u00e1rn\u00ed RNN), kter\u00e9 se pou\u017e\u00edvaj\u00ed v b\u011b\u017en\u00e9m AI modelu Qwen pro uchov\u00e1n\u00ed zm\u011bn\u011bn\u00e9ho stavu p\u0159i pr\u00e1ci s velk\u00fdmi kontextov\u00fdmi okny. Fixac\u00ed stavu br\u00e1ny a aplikac\u00ed unitarizovan\u00e9ho k\u00f3dov\u00e1n\u00ed (one-hot) ka\u017ed\u00fd po\u017eadavek na z\u00e1pis p\u0159episuje p\u0159esn\u011b jeden \u0159\u00e1dek matice stavu, co\u017e umo\u017e\u0148uje pou\u017e\u00edvat stav DeltaNet jako adresovatelnou pam\u011b\u0165 pevn\u00e9 velikosti a vyhnout se line\u00e1rn\u00edmu r\u016fstu KV cache (Key-Value) p\u0159i zvy\u0161ov\u00e1n\u00ed d\u00e9lky kontextu. V takov\u00e9 pam\u011bti jsou um\u00edst\u011bny registry CPU, data a vykon\u00e1van\u00fd strojov\u00fd k\u00f3d. <\/p>\n
<\/center><\/p>\n
Instrukce procesoru, jako je \u010dten\u00ed registr\u016f, aritmetick\u00e9 operace a podm\u00ednkov\u00e9 p\u0159echody, jsou kompilov\u00e1ny p\u0159\u00edmo do vah vrstev transform\u00e1toru. Ka\u017ed\u00e9 instrukci jsou p\u0159i\u0159azeny speci\u00e1ln\u00ed tokeny (ADDR, VAL, STEP), kter\u00e9 vrstvy AI modelu interpretuj\u00ed jako logiku pr\u00e1ce CPU. K inicializaci CPU a na\u010dten\u00ed firmwaru se pou\u017e\u00edv\u00e1 mal\u00fd extern\u00ed ovlada\u010d RvChat, ale samotn\u00fd strojov\u00fd k\u00f3d je zpracov\u00e1v\u00e1n pln\u011b uvnit\u0159 vah modelu.<\/p>\n
Mezi omezen\u00edmi projektu se zmi\u0148uje p\u0159ibli\u017en\u011b dvojn\u00e1sobn\u00e9 zv\u00fd\u0161en\u00ed velikosti (18,8 GB) ve srovn\u00e1n\u00ed se z\u00e1kladn\u00edm modelem Qwen p\u0159i uchov\u00e1v\u00e1n\u00ed stavov\u00fdch sn\u00edmk\u016f (checkpoint), co\u017e vypl\u00fdv\u00e1 p\u0159edev\u0161\u00edm z dodate\u010dn\u00e9ho vypln\u011bn\u00ed dal\u0161\u00edch vrstev. Tak\u00e9 se uv\u00e1d\u00ed relativn\u011b n\u00edzk\u00e1 rychlost v\u00fdpo\u010dt\u016f (p\u0159ibli\u017en\u011b 3200 token\u016f za sekundu na GPU L4), kv\u016fli \u010demu\u017e mohou slo\u017eit\u00e9 v\u00fdpo\u010dty trvat sekundy, nikoli milisekundy. <\/p>\n
D\u00e1le byl realizov\u00e1n variant AI-modelu s experiment\u00e1ln\u00edm souborov\u00fdm syst\u00e9mem, kter\u00fd umo\u017e\u0148uje uchov\u00e1vat uvnit\u0159 vah AI-modelu a\u017e 512 soubor\u016f s 53 tis\u00edci tokeny textu. Stejn\u011b jako v implementaci opera\u010dn\u00ed pam\u011bti je uchov\u00e1v\u00e1n\u00ed organizov\u00e1no v\u00e1zan\u011b na stavy DeltaNet, ale li\u0161\u00ed se t\u00edm, \u017ee FS zaji\u0161\u0165uje trval\u00e9 ulo\u017een\u00ed, kter\u00e9 neztr\u00e1c\u00ed stav mezi relacemi.<\/p>\n
V\u00fdvoj Qwenputer tak\u00e9 poslou\u017eil jako z\u00e1klad pro zkoum\u00e1n\u00ed nov\u00fdch potenci\u00e1ln\u00edch bezpe\u010dnostn\u00edch probl\u00e9m\u016f v AI. Na p\u0159\u00edkladu Qwenputer je uk\u00e1z\u00e1no, jak speci\u00e1ln\u011b modifikovan\u00fd AI-model m\u016f\u017ee skr\u00fdvat sv\u00e9 \u00favahy ve stavu RNN a obch\u00e1zet mechanismy monitorov\u00e1n\u00ed \u00favah modelu (CoT, Chain-of-Thought). Tak\u00e9 byla demonstrov\u00e1na \nmo\u017enost vytv\u00e1\u0159en\u00ed mechanism\u016f kontroly a monitorov\u00e1n\u00ed bezpe\u010dnosti, zabudovan\u00fdch do vah AI-modelu, kter\u00e9 sama model nem\u016f\u017ee vypnout nebo zm\u011bnit. \n Zdroj: 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