Heano maitiro ekushanda ne gpt-oss-20b munharaunda: chii chitsva, kuita, uye maitiro ekuchiyedza.

Kugadziridza kwekupedzisira: 28/08/2025

  • gpt-oss-20b inosvika seyakavhurika-huremu modhi ine kuuraya kwenzvimbo uye kureba mamiriro (kusvika 131.072 tokens).
  • Yakagadzirirwa NVIDIA RTX: Yakashumwa kumhanya kusvika ku256 t/s; VRAM inotora kuti ichengetedze kushanda.
  • Zviri nyore kushandisa neOllama uye dzimwe nzira dzakadai sellama.cpp, GGML, uye Microsoft AI Foundry Local.
  • Inowanikwawo muIntel AI Playground 2.6.0, ine masisitimu akagadziridzwa uye yakagadziridzwa nharaunda manejimendi.
gpt-oss-20b panzvimbo

Kuuya kwe gpt-oss-20b ye kushandiswa kwenzvimbo inounza ine simba yekufunga modhi inomhanya yakananga paPC kune vakawanda vashandisi. Kusundira uku, kunoenderana ne Optimization yeNVIDIA RTX GPUs, inovhura musuwo wekuda kufambiswa kwebasa pasina kuvimba negore.

Chinangwa chakajeka: kupa yakavhurika huremu ine mamiriro akareba kwazvo kumabasa akaoma akadai sekutsvaga kwepamusoro, kutsvaga, rubatsiro rwekodhi kana hurukuro refu, kuisa mberi wega uye kudzora mutengo pakushanda munharaunda.

Chii chinopihwa negpt-oss-20b kana uchimhanya munharaunda?

Kuitwa kwenzvimbo kweakavhurika uremu GPT modhi

Iyo gpt-oss mhuri inotanga nemhando dze vhura uremu yakagadzirirwa kuve nyore kubatanidzwa mune yako mhinduro. Zvakananga, gpt-oss-20b Iyo inomira kunze kwekuenzanisa kugona kufunga uye inonzwisisika hardware zvinodiwa kune desktop PC.

Chinhu chinosiyanisa ndeche hwindo rechirevo chakawedzerwa, nerutsigiro rwekusvika 131.072 tokens mune gpt-oss renji. Kureba uku kunofambisa hurukuro refu, kuongororwa kwemagwaro akakura kana cheni dzakadzika dzepfungwa pasina kucheka kana kupatsanuka.

Exclusive content - Click Here  Yakawanda sei RAM Inoita Windows 10 Inoda?

Kuenzaniswa neakavharwa mamodheru, iyo yakavhurika-huremu chikumbiro chinoisa pamberi kubatanidzwa kushanduka mumashandisirwo: kubva vabatsiri vane maturusi (agents) kunyange plugins ye tsvagiridzo, kutsvaga kwewebhu uye kuronga, vese vachitora mukana wekufungidzira munharaunda.

Nenzira inoshanda, iyo package ye gpt-oss:20b iri kutenderera 13 GB yakaiswa munzvimbo dzakakurumbira dzekumhanya. Izvi zvinogadzirisa toni yezviwanikwa zvinodiwa uye zvinobatsira kuyera iyo VRAM kuchengetedza kushanda pasina zvipingamupinyi.

Kune zvakare yakakura musiyano (gpt-oss-120b), yakagadzirirwa mamiriro ane yakawanda yakawanda graphic resources. Kune akawanda maPC, zvisinei, iyo 20B Ndiyo inonyanya kuitika yekutanga nzvimbo nekuda kwehukama hwayo pakati pekumhanya, ndangariro uye mhando.

Kugadziridza RTX: Kumhanya, Mukati, uye VRAM

Zvishandiso zvekushandisa gpt-oss 20b munharaunda

Kuchinjira maGPT-OSS modhi kune ecosystem NVIDIA RTX inobvumira mitero yechizvarwa chepamusoro. Mumidziyo yepamusoro, mapeaks anosvika 256 tokens/sekondi nekugadzirisa kwakakodzera, kutora mukana weiyo chaiyo optimizations uye nemazvo senge MXFP4.

Mhedzisiro inoenderana nekadhi, mamiriro, uye kugadzirisa. Mukuedzwa ne a RTX 5080, gpt-oss 20b yakasvika kumativi 128 t/s ine zvirimo (≈8k). Nekuwedzera 16k hwindo uye nekumanikidza imwe yemutoro muhurongwa RAM, mwero wakadonha kusvika ~50,5 t/s, neGPU ichiita basa rakawanda.

Exclusive content - Click Here  Maitiro ekugadzirisa Rfc

Chidzidzo chacho chakajeka: the VRAM mitemo. Munzvimbo yeAI, a RTX 3090 ine imwe ndangariro Inogona kuita zvirinani pane nyowani GPU asi iine VRAM shoma, nekuti inodzivirira kufashukira kune system memory uye kuwedzera kupindira kweCPU.

Kune gpt-oss-20b, zviri nyore kutora saizi yemuenzaniso sereferensi: nezve 13 GB imwe nzvimbo ye KV cache uye mabasa akasimba. Semutungamiri wekukurumidza, zvinokurudzirwa kuva nazvo 16 GB yeVRAM zvishoma uye chinangwa 24 GB kana mamiriro akareba kana mitoro yakasimudzwa ichitarisirwa.

Avo vari kutsvaga kusvina hardware vanogona kuongorora kunyatsoita nemazvo (senge MXFP4), gadzirisa kureba kwechinyorwa kana kutendeukira kune akawanda-GPU zvigadziriso kana zvichikwanisika, uchigara uchichengeta chinangwa che dzivisa swaps kuenda ku RAM.

Kuiswa uye kushandiswa: Ollama nedzimwe nzira

GPT-OSS Kuita paRTX GPUs

Kuedza modhi nenzira iri nyore, Ollama inopa chiitiko chakananga paRTX-powered PCs: Inokubvumidza kudhawunirodha, kumhanya, uye kutaura neGPT-OSS-20B pasina magadzirirwo akaomarara., kuwedzera pakutsigira maPDF, mafaera emavara, mapikicha ekusimudzira, uye kugadzirisa mamiriro.

Kune zvakare dzimwe nzira dzevashandisi vepamberi, semuenzaniso Isa LLM paWindows 11. Maframeworks akafanana call.cpp uye nyora maraibhurari GGML dzakagadzirirwa RTX, nekuedza kwemazuva ano mukati kuderedza CPU mutoro uye tora mukana CUDA Grafu. Mukufanana, Microsoft AI Foundry Local (mukutarisa) Batanidza modhi kuburikidza neCLI, SDK kana APIs neCUDA uye TensorRT kukwidziridza.

Exclusive content - Click Here  Maitiro ekushandisa Autoruns kubvisa zvirongwa zvinotanga otomatiki pasina mvumo

Mu ecosystem yezvishandiso, Intel AI Nzvimbo Yekutamba 2.6.0 yakabatanidza gpt-oss-20b pakati pesarudzo dzayoIyo yekuvandudza inowedzera yakanaka-grained vhezheni kutonga kune backends uye kudzokorora kune masisitimu akadai OpenVINO, ComfyUI y call.cpp (nerutsigiro rwe rinoputika uye kugadzirisa mamiriro ezvinhu), kufambisa nzvimbo dzakagadzikana dzemunharaunda.

Segwara rekutanga, tarisa iyo Inowanikwa VRAM, dhawunirodha mhando yemhando inoenderana neGPU yako, simbisa iyo token velocity nezvinomiririra zvinokurudzira uye inogadzirisa iyo hwindo remukati kuchengetedza mutoro wese pane graphics kadhi.

Nezvimedu izvi, zvinokwanisika kuvaka vabatsiri ve kutsvaga nekuongorora, zvishandiso zve kuongorora kana zvinotsigira zve programming iyo inomhanya zvachose pakombuta, ichichengetedza data changamire.

Iko kusanganiswa kwe gpt-oss-20b neRTX kukwidziridza, kungwarira kweVRAM manejimendi, uye maturusi akaita seOllama, llama.cpp, kana AI Playground inosimbisa sarudzo yakakura yekumhanyisa kufunga AI munharaunda; nzira inoenzanisa kuita, mutengo, uye kuvanzika pasina kuvimba nemabasa ekunze.

gpt-oss-120b
Nyaya inoenderana:
OpenAI inoburitsa gpt-oss-120b: yakanyanya kuvhurika maremu modhi kusvika parizvino.