Kharashka kaliya 300,000? Dhisida AI supercomputer shakhsi ah oo leh 4 Mac Studio 512GB, hagaha dejinta Kimi-K2.5 oo leh malaayiin parametir
Kharashka kaliya 300,000? Dhisida AI supercomputer shakhsi ah oo leh 4 Mac Studio 512GB, hagaha dejinta Kimi-K2.5 oo leh malaayiin parametir
Xilligan xawaaraha weyn ee moodooyinka waaweyn, dhammaanteen waxaan leenahay riyo: in aan si maxalli ah u orodsiino moodal malaayiin parametir ah oo la mid ah heerka GPT-5. Laakiin xaqiiqadu waa mid adag, moodooyinka malaayiin parametir ah xitaa haddii la qiyaaso 4-bit, waxay u baahan yihiin xusuus badan. H100, B200 waa qaali, maxaa yeelay ma awoodno in aan iibsanno?
Maanta JamePeng wuxuu na tusinayaa sida aan u dhisi karno supercomputer AI maxalli ah oo leh 4 Mac Studio oo buuxda, iyadoo la adeegsanayo EXO+MLX iyo Thunderbolt 5, si aan u helno 2TB xusuus mideysan! Ujeedadu waa hal: in aan si maxalli ah u orodsiino Kimi-K2.5 moodalka malaayiin parametir ah.
Maxaa loogu baahan yahay in la sameeyo?
Ma aha oo kaliya in la helo waxyaabo xiiso leh, laakiin sidoo kale waa arrin ku saabsan asturnaanta xogta iyo xakameyn buuxda oo maxalli ah.
Hubka ugu muhiimsan waa EXO (GitHub: exo-explore/exo), waxay taageertaa RDMA (Remote Direct Memory Access), waxayna isku dari kartaa xusuusta mideysan ee 4 Mac si ay u noqoto biyo-xireen weyn oo xusuus ah.
Liiska qalabka: 4 Mac Studio (M3 Ultra, nooca 512GB), xusuusta guud waxay ku dhowdahay 2TB, isku xirka waxaa lagu sameeyaa Thunderbolt 5 (120Gbps bandwidth), nidaamku wuxuu u baahan yahay macOS Tahoe 26.2 ama nooc ka dambeeya.
Tallaabada 1: Dhaqangelinta taageerada RDMA
Hawlgallada ku samee Mac kasta:
- Jooji Mac, geli habka soo kabashada (riix badhanka korontada, dooro "Options" > "Continue")
- Fur Terminal, orod: bputil -a rdma
- Dib u bilow Mac
- Hubi: systemprofiler SPThunderboltDataType si aad u hubiso in RDMA la dhaqan geliyay
Tallaabada 2: Ku rakib EXO
Rakibaadda macOS App: Ka soo dejiso GitHub EXO-version.dmg, rakib oo orod. Fur Dashboard si aad ugu darto IP-yada Mac kale.
Rakibaadda koodhka ilaha:
- Ku rakib Homebrew
- git clone https://github.com/exo-explore/exo.git
- pip install -e .
- exo start
Tallaabada 3: Isku xirka jireed iyo topology
Ha isticmaalin Wi-Fi si aad u dhisto shabakad! Xitaa Wi-Fi 7 ma shaqaynayso. Falanqaynta moodooyinka malaayiin parametir ah waxay aad ugu xasaasisan tahay bandwidth. Fadlan isticmaal fiilo Thunderbolt 5, hal Mac u samee sida node-ka ugu weyn (Master), saddexda kale u samee sida nodes-ka shaqada (Worker). Waxaa lagu talinayaa topology xiddig ama isku xidhka silsiladda.
Dashboard-ka EXO, waa inaad aragtaa 4 qalab oo dhammaantood online ah, xusuusta guudna waxay muujinaysaa 2048 GB.
Tallaabada 4: Soo dejinta iyo orodka nooca bulshada ee MLX Kimi-K2.5
pip install huggingfacehub huggingface-cli download mlx-community/Kimi-K2.5 --local-dir ./models/mlx-community/Kimi-K2.5 2. Bilow mashiinka falanqaynta:
exo run --model ./models/mlx-community/Kimi-K2.5 --quant 4 --shards auto --engine mlx Falanqaynta amarka:
- --model: tilmaamaya galka moodalka
- --quant 4: isticmaalka 4-bit qiyaasta si loo yareeyo isticmaalka xusuusta
- --shards auto: EXO si toos ah u kala qaybinaya moodalka
- --engine mlx: wacitaanka 76-core GPU iyo Neural Engine ee M3 Ultra si loo sameeyo falanqayn
Natiijada ugu dambeysa iyo tijaabada
Marka terminal-ka uu muujiyo Ready, waxaad haysataa supercomputer AI oo kuu gaar ah.
Marxaladda hore (Prefill): Fanka 4 Mac ayaa bilaabaya in ay si tartiib ah u xawaareeyaan (waxaa ka faa'iideysanaya waxtarka M3 Ultra, ma noqon doonto mid duulaysa).
Marxaladda abuurista (Generation): Token-ka hal midba midka kale ayaa soo baxaya.
Xawaaraha: In kasta oo aanay la tartamin H100 cluster, haddana waxaa ka faa'iideysanaya RDMA ee Thunderbolt 5, xawaaraha abuurista Token-ka wuxuu gaari karaa 17-28 tokens/s. Tani waxay ku habboon tahay moodal malaayiin parametir ah, gabi ahaanba waa la falanqeyn karaa!
Gunaanad
Nidaamkan gabi ahaanba ma ahan mid raqiis ah, laakiin wuxuu caddeynayaa in dadaalka Apple Silicon + bulshada furan uu keeni doono mustaqbalka AI ee aan dhexeeyn. Ma u baahnin in aan xogta u dirno shirkadaha daruuraha, iyadoo la adeegsanayo qalabka haysanno, waxaan dhisi karnaa kooxo falanqayn oo xoog leh.

