Kuphunzitsa ndi kuyendetsa PyTorch mafano pa NVIDIA Tesla P40 (24 GB VRAM, Pascal kapangidwe), ndi kulowa kwathunthu root. GPU ma seva si ovomerezeka pano; tiuzeni kuti adziwidwe pamene iwo ndi.
$ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
# Mfano wa kukhazikitsa lamulo
PyTorch ndi chida chogwiritsa ntchito kwambiri chophunzira. GPU server imayendetsa PyTorch kuphunzitsa ndi kuzindikira pa GPU, yomwe ndi yofulumira kwambiri kuposa CPU kwa mamodeli ambiri.
Mumayika NVIDIA driver, CUDA ndi PyTorch build yomwe mukufuna. Pezani kuti PyTorch build yanu ikugwira ntchito ndi Pascal GPUs (compute capability 6.1).
Malo a mapangidwe azing'ono ndi apakati, LoRA fine-tuning ndi kukula kwa mayunitsi otsika.
Kuchita Jupyter notebooks pa seva kwa ntchito yolumikizana pa GPU.
Configure chilengedwe mofanana monga zosowa wanu ntchito.
PyTorch akhoza kugwiritsa ntchito NVIDIA GPU kuti afulumizitse ntchito yake. VPS.org GPU ma seva si angathe kulandira; tiuzeni kuti mudziwe pamene iwo ndi. PyTorch ndi maphunziro / fine-tuning workload. Plan for long running jobs: snapshot your VPS regularly and keep a copy of trained weights outside VPS.org.
GPU ma seva si opezeka kuti aperekebe. Tsopano mungagwiritse ntchito PyTorch pa CPU yachikhalidwe, komwe imagwira ntchito mofulumira kuposa pa GPU: SSH mu ndi kuyendetsa pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124.
Kuphunzitsa VRAM kumadziwika ndi lamulo la optimization komanso kuyambitsa. Kuthamanga kwa 7B LLM kumafunikira pafupifupi 24 mpaka 48 GB; LoRA / QLoRA imagwirizana ndi 8 mpaka 16 GB.
Kugulitsa kwa GPU kudzatulutsidwa pamene ma seva a GPU angagwiritsidwe ntchito. Tiuzeni ngati mukufuna kumvetsetsa pamene izi zikuchitika.
Yai. Kawirikawiri VPS.org server amachokera ndi wonse root kulowa, kotero inu mukhoza kuyendetsa ena zipangizo mbali PyTorch m'ma RAM ndi kusungirako server.
Gawo lililonse VPS.org wothandizira amachokera ndi wonse root SSH kulowa, kotero inu mukhoza kukhazikitsa oyendetsa, mabuku ndi PyTorch koma inu muyenera.
Sichoncho. Palibe wothandizira amafika ndi CUDA runtime kapena NVIDIA woyendetsa preinstalled, ndi GPU othandizira sizilipo kuti aperekebe. Mumakhazikitsa PyTorch zojambula zofunikira nokha.
Ndikoyenera. VPS.org ma seva ndi olimba olimba, osati opanda pake: PyTorch mapangidwe, zosankha ndi deta zimakhazikika pa SSD yosungira server mpaka mutachotsa.
Kusunga ntchito deta pa SSD seva kwa kulowa mosavuta pamene PyTorch kunyamula, ndi kukopera kumaliza artifacts (mawuwo, chiyambi, embeddings) ku osungirako kunja VPS.org; snapshots ndi zosunga zobwezeretsera azikhala pamodzi osungirako gulu.
Mukhoza kuwonjezera iliyonse VPS.org wothandizira kuti malire lalikulu kuchokera ku gulu la kuwongolera; resize imayambanso ndi PyTorch kukhazikitsa wanu amapita. GPU othandizira si opezekabe.
Yes. Backups ndi wolipira add-on (tsiku ndi tsiku + 30% ya mtengo wa ndondomeko, anasunga masiku 7; tsiku ndi tsiku + 20%, anasunga masiku 28) ndi snapshots ndi ufulu, amene ndi zothandiza pamaso PyTorch m'mbuyo m'mbuyo.
Mutha kulumikizana ndi thandizo m'masiku 30 a kugula kwanu koyamba kuti mupemphe kuti mubwezere ndalama. GPU servers sizikugulitsidwabe, chifukwa chake palibe GPU order.
Kutha kwa GPU ndi kocheperako. Tiuzeni za ntchito yanu ndipo tidzakuyankha ndi kuchuluka kwa ndalama zomwe zilipo.