GPU Droplet’te Nemotron 3 Modelleri Nas\u0131l \u00c7al\u0131\u015ft\u0131r\u0131l\u0131r?<\/title><br \/>\n<body><\/p>\n<p>\n A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.\n<\/p>\n<h2>Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?<\/h2>\n<p>\n Yapay zeka d\u00fcnyas\u0131 h\u0131zla geli\u015firken, \u00f6zellikle b\u00fcy\u00fck dil modelleri (LLM’ler) alan\u0131nda \u00e7\u0131\u011f\u0131r a\u00e7an yeniliklere tan\u0131k oluyoruz. NVIDIA’n\u0131n Nemotron 3 serisi, a\u00e7\u0131k kaynak toplulu\u011funa sunulan g\u00fc\u00e7l\u00fc ve yetenekli modellerle bu de\u011fi\u015fimin \u00f6nc\u00fclerinden biri haline geldi. Peki, bu modelleri neden bir “GPU Droplet” \u00fczerinde \u00e7al\u0131\u015ft\u0131rmay\u0131 d\u00fc\u015f\u00fcnmeliyiz? Bu sorunun cevab\u0131, hem performans hem de maliyet etkinli\u011fi a\u00e7\u0131s\u0131ndan kritik \u00f6neme sahip. Nemotron 3 modelleri, do\u011fal dil anlama, metin \u00fcretimi, \u00f6zetleme ve \u00e7eviri gibi karma\u015f\u0131k g\u00f6revleri ba\u015far\u0131yla yerine getirebilen, milyarlarca parametreye sahip modellerdir. Bu modellerin tam potansiyelini ortaya \u00e7\u0131karmak i\u00e7in y\u00fcksek hesaplama g\u00fcc\u00fcne ve geni\u015f belle\u011fe sahip grafik i\u015flem birimlerine (GPU’lar) ihtiya\u00e7 duyulur. Geleneksel CPU’lar, bu t\u00fcr paralel i\u015flemleri verimli bir \u015fekilde ger\u00e7ekle\u015ftiremezken, GPU’lar binlerce \u00e7ekirde\u011fi sayesinde bu i\u015f y\u00fck\u00fcn\u00fc kolayca kald\u0131rabilir.\n<\/p>\n<p>\n Bir “GPU Droplet”, DigitalOcean gibi bulut sa\u011flay\u0131c\u0131lar\u0131n\u0131n sundu\u011fu, \u00f6nceden yap\u0131land\u0131r\u0131lm\u0131\u015f bir sanal sunucu hizmetidir ve i\u00e7inde g\u00fc\u00e7l\u00fc bir GPU bar\u0131nd\u0131r\u0131r. Bu, fiziksel donan\u0131m sat\u0131n alma ve bak\u0131m maliyetlerinden ka\u00e7\u0131narak, ihtiyac\u0131n\u0131z oldu\u011funda y\u00fcksek performansl\u0131 bir GPU’ya eri\u015fmenizi sa\u011flar. \u00d6zellikle k\u00fc\u00e7\u00fck ve orta \u00f6l\u00e7ekli geli\u015ftiriciler, ara\u015ft\u0131rmac\u0131lar veya start-up’lar i\u00e7in bu, oyunun kurallar\u0131n\u0131 de\u011fi\u015ftiren bir \u00e7\u00f6z\u00fcmd\u00fcr. Kendi sunucunuzu kurmak ve y\u00f6netmek yerine, birka\u00e7 t\u0131klamayla dakikalar i\u00e7inde kullan\u0131ma haz\u0131r bir yapay zeka geli\u015ftirme ortam\u0131na sahip olabilirsiniz. Ayr\u0131ca, GPU Droplet’ler \u00f6l\u00e7eklenebilirlik avantaj\u0131 sunar; projenizin gereksinimleri artt\u0131\u011f\u0131nda daha g\u00fc\u00e7l\u00fc bir Droplet’e ge\u00e7i\u015f yapabilir veya ihtiyac\u0131n\u0131z kalmad\u0131\u011f\u0131nda kolayca kapatabilirsiniz. Bu esneklik, maliyetleri optimize etmenize ve sadece kulland\u0131\u011f\u0131n\u0131z kadar \u00f6deme yapman\u0131za olanak tan\u0131r.\n<\/p>\n<p>\n Nemotron 3 gibi a\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 modeller, kullan\u0131c\u0131lara tam \u015feffafl\u0131k ve \u00f6zelle\u015ftirme imkan\u0131 sunar. Bu, modelin i\u00e7 i\u015fleyi\u015fini anlama, kendi veri k\u00fcmelerinizle ince ayar yapma (fine-tuning) veya belirli g\u00f6revler i\u00e7in optimize etme f\u0131rsat\u0131 demektir. Bir GPU Droplet \u00fczerinde bu modelleri \u00e7al\u0131\u015ft\u0131rmak, bu \u00f6zg\u00fcrl\u00fc\u011f\u00fc pratik bir \u015fekilde hayata ge\u00e7irmenizi sa\u011flar. Geli\u015ftiriciler, modelin \u00e7\u0131kt\u0131s\u0131n\u0131 do\u011frudan etkileyebilir, performans\u0131n\u0131 izleyebilir ve hatta yeni mimarilerle deneyler yapabilirler. \u00d6zellikle gizlilik endi\u015feleri olan veya hassas verilerle \u00e7al\u0131\u015fan \u015firketler i\u00e7in, kendi kontrol\u00fcn\u00fcz alt\u0131ndaki bir bulut ortam\u0131nda a\u00e7\u0131k kaynakl\u0131 modelleri kullanmak, verilerin \u00fc\u00e7\u00fcnc\u00fc taraf API’lara g\u00f6nderilmesi riskini ortadan kald\u0131r\u0131r. Bu kombinasyon, hem teknik kapasiteyi art\u0131r\u0131r hem de yapay zeka projelerinin geli\u015ftirme d\u00f6ng\u00fcs\u00fcn\u00fc h\u0131zland\u0131r\u0131r, b\u00f6ylece yenilik\u00e7i uygulamalar\u0131n \u00f6n\u00fcn\u00fc a\u00e7ar. Dolay\u0131s\u0131yla, Nemotron 3 ve GPU Droplet birle\u015fimi, modern yapay zeka uygulamalar\u0131 i\u00e7in g\u00fc\u00e7l\u00fc, esnek ve maliyet etkin bir temel olu\u015fturur.\n<\/p>\n<h2>GPU Droplet Se\u00e7imi ve Temel Kurulum Ad\u0131mlar\u0131: Hangi Droplet Sizin \u0130\u00e7in \u0130deal?<\/h2>\n<p>\n Nemotron 3 gibi b\u00fcy\u00fck dil modellerini verimli bir \u015fekilde \u00e7al\u0131\u015ft\u0131rmak i\u00e7in do\u011fru GPU Droplet’i se\u00e7mek, projenizin ba\u015far\u0131s\u0131 i\u00e7in kritik \u00f6neme sahiptir. Piyasada farkl\u0131 \u00f6zelliklere ve fiyatland\u0131rmalara sahip bir\u00e7ok GPU Droplet se\u00e7ene\u011fi bulunmaktad\u0131r. Peki, bu se\u00e7enekler aras\u0131ndan size en uygun olan\u0131 nas\u0131l belirleyeceksiniz? \u0130lk olarak, modelin bellek (VRAM) gereksinimlerini g\u00f6z \u00f6n\u00fcnde bulundurmal\u0131s\u0131n\u0131z. Nemotron 3 modelleri, boyutlar\u0131na ba\u011fl\u0131 olarak \u00f6nemli miktarda VRAM’e ihtiya\u00e7 duyabilir. \u00d6rne\u011fin, 8 milyar parametreli bir model i\u00e7in en az 16 GB VRAM \u00f6nerilirken, daha b\u00fcy\u00fck modeller i\u00e7in 24 GB veya daha fazlas\u0131 gerekebilir. Bu nedenle, se\u00e7ti\u011finiz Droplet’in GPU’sunun yeterli VRAM’e sahip oldu\u011fundan emin olmal\u0131s\u0131n\u0131z. NVIDIA’n\u0131n A100 veya H100 gibi y\u00fcksek performansl\u0131 GPU’lar sunan Droplet’ler, daha b\u00fcy\u00fck modeller veya daha h\u0131zl\u0131 \u00e7\u0131kar\u0131m (inference) s\u00fcre\u00e7leri i\u00e7in idealdir. Ancak, daha k\u00fc\u00e7\u00fck modeller veya b\u00fct\u00e7e k\u0131s\u0131tlamalar\u0131 olan projeler i\u00e7in RTX serisi GPU’lar da uygun bir ba\u015flang\u0131\u00e7 noktas\u0131 olabilir.\n<\/p>\n<p>\n Droplet se\u00e7iminde dikkate alman\u0131z gereken bir di\u011fer \u00f6nemli fakt\u00f6r, CPU \u00e7ekirdek say\u0131s\u0131 ve RAM miktar\u0131d\u0131r. GPU, paralel hesaplamalar i\u00e7in ana i\u015fi yaparken, CPU ve sistem RAM’i model y\u00fckleme, veri \u00f6n i\u015fleme ve di\u011fer yard\u0131mc\u0131 g\u00f6revler i\u00e7in hala \u00f6nemlidir. Genellikle, en az 4-8 CPU \u00e7ekirde\u011fi ve 32 GB RAM, sorunsuz bir deneyim i\u00e7in iyi bir denge sunar. Bulut sa\u011flay\u0131c\u0131n\u0131z\u0131n (\u00f6rne\u011fin DigitalOcean, AWS, Google Cloud) sundu\u011fu farkl\u0131 GPU Droplet planlar\u0131n\u0131 kar\u015f\u0131la\u015ft\u0131r\u0131rken, ayl\u0131k veya saatlik maliyetleri, veri transfer \u00fccretlerini ve b\u00f6lge se\u00e7eneklerini de de\u011ferlendirmelisiniz. Co\u011frafi olarak size veya hedef kitlenize yak\u0131n bir b\u00f6lge se\u00e7mek, gecikmeyi azaltarak performans\u0131 art\u0131rabilir. Bir kez Droplet’inizi se\u00e7ip olu\u015fturduktan sonra, temel kurulum ad\u0131mlar\u0131na ge\u00e7ebiliriz. Genellikle bu s\u00fcre\u00e7 SSH (Secure Shell) \u00fczerinden sunucuya ba\u011flanarak ba\u015flar.\n<\/p>\n<pre><code class=\"language-bash\">\nssh root@sizin_droplet_ip_adresi\n<\/code><\/pre>\n<p>\n Bu komutla sunucunuza g\u00fcvenli bir \u015fekilde ba\u011flanabilirsiniz. \u0130lk ba\u011flant\u0131da genellikle parolan\u0131z\u0131 girmeniz istenir. Ba\u011fland\u0131ktan sonra, i\u015fletim sisteminizin g\u00fcncel oldu\u011fundan emin olmak \u00f6nemlidir. Ubuntu gibi Debian tabanl\u0131 sistemlerde bu i\u015flemi a\u015fa\u011f\u0131daki komutlarla yapabilirsiniz:\n<\/p>\n<pre><code class=\"language-bash\">\nsudo apt update\nsudo apt upgrade -y\n<\/code><\/pre>\n<p>\n Bu komutlar, sisteminizdeki paket listelerini g\u00fcnceller ve mevcut t\u00fcm paketleri en son s\u00fcr\u00fcmlerine y\u00fckseltir. Ard\u0131ndan, NVIDIA s\u00fcr\u00fcc\u00fclerini ve CUDA Toolkit’i kurman\u0131z gerekecektir. Bu ad\u0131mlar, GPU’nuzun yaz\u0131l\u0131mlarla do\u011fru \u015fekilde ileti\u015fim kurmas\u0131n\u0131 sa\u011flar. NVIDIA’n\u0131n web sitesinden g\u00fcncel s\u00fcr\u00fcc\u00fcleri ve CUDA Toolkit’i indirmeniz ve kurulum talimatlar\u0131n\u0131 takip etmeniz en g\u00fcvenli yoldur. Genellikle, CUDA Toolkit kurulumu s\u0131ras\u0131nda s\u00fcr\u00fcc\u00fcler de otomatik olarak kurulur. Ancak bazen manuel kurulum gerekebilir. Bu s\u00fcre\u00e7, komut sat\u0131r\u0131 \u00fczerinden <code class=\"language-\">.deb<\/code> veya <code class=\"language-\">.run<\/code> dosyalar\u0131n\u0131 \u00e7al\u0131\u015ft\u0131rmay\u0131 i\u00e7erir. Kurulum tamamland\u0131ktan sonra, GPU’nuzun do\u011fru \u015fekilde tan\u0131nd\u0131\u011f\u0131n\u0131 ve \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 do\u011frulamak i\u00e7in <code class=\"language-\">nvidia-smi<\/code> komutunu kullanabilirsiniz. Bu komut, GPU’nuzun durumunu, bellek kullan\u0131m\u0131n\u0131 ve \u00e7al\u0131\u015fan s\u00fcre\u00e7leri g\u00f6sterir.\n<\/p>\n<pre><code class=\"language-bash\">\nnvidia-smi\n<\/code><\/pre>\n<p>\n Bu \u00e7\u0131kt\u0131da GPU’nuzun ad\u0131n\u0131, s\u00fcr\u00fcc\u00fc s\u00fcr\u00fcm\u00fcn\u00fc ve mevcut VRAM miktar\u0131n\u0131 g\u00f6rmelisiniz. E\u011fer bu komut d\u00fczg\u00fcn \u00e7al\u0131\u015f\u0131yorsa, temel GPU kurulumunuz ba\u015far\u0131yla tamamlanm\u0131\u015f demektir. Bu a\u015famadan sonra, Nemotron 3 modellerini \u00e7al\u0131\u015ft\u0131rmak i\u00e7in gerekli Python ortam\u0131n\u0131 ve k\u00fct\u00fcphanelerini kurmaya haz\u0131r olacaks\u0131n\u0131z. Unutmay\u0131n, her ad\u0131mda dikkatli olmak ve herhangi bir hata durumunda ilgili dok\u00fcmantasyonlara ba\u015fvurmak, sorunsuz bir kurulum s\u00fcreci i\u00e7in anahtard\u0131r.\n<\/p>\n<div class=\"expert-tip\">\n Uzman \u0130pucu: Droplet’inizi se\u00e7erken, sadece anl\u0131k model boyutunu de\u011fil, gelecekteki olas\u0131 b\u00fcy\u00fcme veya ince ayar (fine-tuning) ihtiya\u00e7lar\u0131n\u0131z\u0131 da g\u00f6z \u00f6n\u00fcnde bulundurun. Biraz daha fazla VRAM’e sahip bir Droplet se\u00e7mek, uzun vadede size esneklik sa\u011flayabilir. Ayr\u0131ca, DigitalOcean gibi platformlarda “snapshot” \u00f6zelli\u011fi ile \u00e7al\u0131\u015fan bir ortam\u0131n yede\u011fini alarak olas\u0131 sorunlara kar\u015f\u0131 haz\u0131rl\u0131kl\u0131 olabilirsiniz.\n<\/div>\n<h2>Gerekli Yaz\u0131l\u0131mlar\u0131n Kurulumu: Yapay Zeka Ortam\u0131n\u0131z\u0131 Haz\u0131rlama<\/h2>\n<p>\n GPU Droplet’inizde temel i\u015fletim sistemi ve NVIDIA s\u00fcr\u00fcc\u00fcleri kurulduktan sonra, Nemotron 3 modellerini \u00e7al\u0131\u015ft\u0131rmak i\u00e7in gerekli olan yapay zeka yaz\u0131l\u0131m ortam\u0131n\u0131 haz\u0131rlama a\u015famas\u0131na ge\u00e7iyoruz. Bu ad\u0131m, Python, gerekli k\u00fct\u00fcphaneler (PyTorch veya TensorFlow) ve Hugging Face Transformers gibi ara\u00e7lar\u0131n kurulumunu i\u00e7erir. Do\u011fru bir kurulum, modellerin sorunsuz ve verimli \u00e7al\u0131\u015fmas\u0131n\u0131 sa\u011flayacakt\u0131r. \u0130lk olarak, modern yapay zeka uygulamalar\u0131n\u0131n temelini olu\u015fturan Python’\u0131 kurmam\u0131z gerekiyor. Genellikle Droplet’lerde Python zaten y\u00fckl\u00fc gelir, ancak g\u00fcncel bir s\u00fcr\u00fcm (genellikle 3.8 ve \u00fczeri) kullanmak en iyisidir. Python’\u0131n farkl\u0131 s\u00fcr\u00fcmlerini y\u00f6netmek i\u00e7in <code class=\"language-\">pyenv<\/code> veya <code class=\"language-\">conda<\/code> gibi ara\u00e7lar\u0131 kullanmak, projenin ba\u011f\u0131ml\u0131l\u0131klar\u0131n\u0131 izole etmek a\u00e7\u0131s\u0131ndan faydal\u0131d\u0131r. Bu rehberde, <code class=\"language-\">venv<\/code> mod\u00fcl\u00fc ile basit bir sanal ortam olu\u015fturmay\u0131 tercih edece\u011fiz.\n<\/p>\n<pre><code class=\"language-bash\">\nsudo apt install python3-pip python3.10-venv -y # Python 3.10 ve venv kurar\npython3.10 -m venv nemotron_env # nemotron_env ad\u0131nda bir sanal ortam olu\u015fturur\nsource nemotron_env\/bin\/activate # Sanal ortam\u0131 etkinle\u015ftirir\n<\/code><\/pre>\n<p>\n Sanal ortam\u0131 etkinle\u015ftirdikten sonra, t\u00fcm Python paketleri bu ortama y\u00fcklenecek ve sistem genelindeki Python kurulumunuzu etkilemeyecektir. Bu, ba\u011f\u0131ml\u0131l\u0131k \u00e7ak\u0131\u015fmalar\u0131n\u0131 \u00f6nlemek i\u00e7in iyi bir uygulamad\u0131r. \u015eimdi s\u0131ra, derin \u00f6\u011frenme \u00e7er\u00e7evelerini kurmaya geldi. Nemotron 3 modelleri genellikle PyTorch ile uyumludur, bu y\u00fczden PyTorch kurulumuna odaklanaca\u011f\u0131z. CUDA deste\u011fiyle PyTorch kurmak, GPU’nuzun t\u00fcm g\u00fcc\u00fcn\u00fc kullanman\u0131z\u0131 sa\u011flar. PyTorch’un resmi web sitesinden (pytorch.org) g\u00fcncel kurulum komutunu alman\u0131z \u00f6nemlidir, \u00e7\u00fcnk\u00fc CUDA s\u00fcr\u00fcmleri ve PyTorch versiyonlar\u0131 aras\u0131nda uyumluluk farkl\u0131l\u0131klar\u0131 olabilir. A\u015fa\u011f\u0131daki \u00f6rnek, genellikle g\u00fcncel bir CUDA s\u00fcr\u00fcm\u00fc i\u00e7in ge\u00e7erlidir:\n<\/p>\n<pre><code class=\"language-bash\">\npip install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/cu118\n<\/code><\/pre>\n<p>\n Buradaki <code class=\"language-\">cu118<\/code>, CUDA 11.8 s\u00fcr\u00fcm\u00fc i\u00e7in derlenmi\u015f PyTorch paketini ifade eder. Kendi CUDA s\u00fcr\u00fcm\u00fcn\u00fcze uygun olan\u0131 se\u00e7ti\u011finizden emin olun. PyTorch’un do\u011fru kuruldu\u011funu ve GPU’nuzu tan\u0131d\u0131\u011f\u0131n\u0131 do\u011frulamak i\u00e7in Python etkile\u015fimli kabu\u011funu a\u00e7\u0131p a\u015fa\u011f\u0131daki komutlar\u0131 \u00e7al\u0131\u015ft\u0131rabilirsiniz:\n<\/p>\n<pre><code class=\"language-python\">\nimport torch\nprint(torch.cuda.is_available())\nprint(torch.cuda.device_count())\nprint(torch.cuda.get_device_name(0))\n<\/code><\/pre>\n<p>\n Bu komutlar\u0131n \u00e7\u0131kt\u0131s\u0131 <code class=\"language-\">True<\/code>, kullan\u0131labilir GPU say\u0131s\u0131 (genellikle 1) ve GPU’nuzun ad\u0131 olmal\u0131d\u0131r. E\u011fer <code class=\"language-\">False<\/code> veya hata al\u0131yorsan\u0131z, PyTorch kurulumunuzda veya NVIDIA s\u00fcr\u00fcc\u00fclerinizde bir sorun olabilir. Son olarak, Nemotron 3 modellerini kolayca y\u00fcklemek ve kullanmak i\u00e7in Hugging Face Transformers k\u00fct\u00fcphanesini kurmam\u0131z gerekiyor. Bu k\u00fct\u00fcphane, \u00f6nceden e\u011fitilmi\u015f modelleri indirme, y\u00fckleme ve \u00e7\u0131kar\u0131m yapma s\u00fcre\u00e7lerini b\u00fcy\u00fck \u00f6l\u00e7\u00fcde basitle\u015ftirir. Ayr\u0131ca, h\u0131zland\u0131rma i\u00e7in <code class=\"language-\">accelerate<\/code> ve <code class=\"language-\">bitsandbytes<\/code> gibi ek k\u00fct\u00fcphaneleri de kurmak, \u00f6zellikle bellek k\u0131s\u0131tl\u0131 ortamlarda veya daha b\u00fcy\u00fck modellerle \u00e7al\u0131\u015f\u0131rken faydal\u0131d\u0131r.\n<\/p>\n<pre><code class=\"language-bash\">\npip install transformers accelerate bitsandbytes\n<\/code><\/pre>\n<p>\n <code class=\"language-\">bitsandbytes<\/code> k\u00fct\u00fcphanesi, modelleri daha d\u00fc\u015f\u00fck hassasiyetle (\u00f6rne\u011fin 8-bit veya 4-bit) y\u00fckleyerek VRAM kullan\u0131m\u0131n\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r ve \u00e7\u0131kar\u0131m h\u0131z\u0131n\u0131 art\u0131rabilir. <code class=\"language-\">accelerate<\/code> ise da\u011f\u0131t\u0131k e\u011fitim ve \u00e7\u0131kar\u0131m i\u00e7in yard\u0131mc\u0131 bir ara\u00e7t\u0131r. Bu ad\u0131mlar\u0131 tamamlad\u0131\u011f\u0131n\u0131zda, GPU Droplet’iniz Nemotron 3 modellerini \u00e7al\u0131\u015ft\u0131rmaya haz\u0131r bir yapay zeka geli\u015ftirme ortam\u0131na d\u00f6n\u00fc\u015fm\u00fc\u015f olacakt\u0131r. T\u00fcm bu k\u00fct\u00fcphanelerin do\u011fru versiyonlarda ve uyumlu bir \u015fekilde kuruldu\u011fundan emin olmak i\u00e7in her ad\u0131mda dikkatli olmak, ba\u015far\u0131l\u0131 bir entegrasyon i\u00e7in anahtard\u0131r. Art\u0131k modelleri indirip kullanmaya ba\u015flayabiliriz!\n<\/p>\n<h2>Nemotron 3 Modelini \u0130ndirme ve Y\u00fckleme: Hugging Face Hub ile Tan\u0131\u015f\u0131n<\/h2>\n<p>\n Yapay zeka modellerini kullanman\u0131n en kolay yollar\u0131ndan biri, Hugging Face Hub’\u0131 kullanmakt\u0131r. Nemotron 3 modelleri de dahil olmak \u00fczere binlerce \u00f6nceden e\u011fitilmi\u015f modelin bar\u0131nd\u0131r\u0131ld\u0131\u011f\u0131 bu platform, geli\u015ftiricilere b\u00fcy\u00fck kolayl\u0131k sa\u011flar. Bu b\u00f6l\u00fcmde, Nemotron 3 modelini Hugging Face Hub’dan nas\u0131l indirece\u011finizi ve PyTorch ile GPU Droplet’inize nas\u0131l y\u00fckleyece\u011finizi ad\u0131m ad\u0131m inceleyece\u011fiz. \u0130lk olarak, hangi Nemotron 3 modelini kullanaca\u011f\u0131n\u0131za karar vermelisiniz. NVIDIA, farkl\u0131 boyutlarda (\u00f6rne\u011fin 8B, 22B, 128B parametreler) ve farkl\u0131 hassasiyetlerde (\u00f6rn. <code class=\"language-\">fp16<\/code>, <code class=\"language-\">bf16<\/code>, <code class=\"language-\">int8<\/code>) Nemotron 3 modelleri sunmaktad\u0131r. Modelin boyutu, ihtiya\u00e7 duyulan VRAM miktar\u0131n\u0131 do\u011frudan etkiler. \u00d6rne\u011fin, 8 milyar parametreli bir model, 22 milyar parametreli bir modelden daha az VRAM gerektirecektir. Droplet’inizin GPU’sunun VRAM kapasitesini g\u00f6z \u00f6n\u00fcnde bulundurarak en uygun modeli se\u00e7melisiniz. Genellikle, daha k\u00fc\u00e7\u00fck modellerle ba\u015flay\u0131p performans ve bellek kullan\u0131m\u0131n\u0131 de\u011ferlendirdikten sonra daha b\u00fcy\u00fck modellere ge\u00e7mek iyi bir stratejidir.\n<\/p>\n<p>\n Modeli indirmek ve y\u00fcklemek i\u00e7in Hugging Face <code class=\"language-\">transformers<\/code> k\u00fct\u00fcphanesini kullanaca\u011f\u0131z. Bu k\u00fct\u00fcphane, modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 otomatik olarak indirir ve GPU’ya y\u00fckler. \u0130\u015fte temel ad\u0131mlar:\n<\/p>\n<pre><code class=\"language-python\">\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport torch\n\n# Kullanmak istedi\u011finiz Nemotron 3 modelinin Hugging Face ID'si\n# \u00d6rnek: \"nvidia\/Nemotron-3-8B-Base\" veya \"nvidia\/Nemotron-3-8B-Instruct\"\nmodel_name = \"nvidia\/Nemotron-3-8B-Instruct\" \n\nprint(f\"{model_name} modelini indiriyor ve y\u00fckl\u00fcyor...\")\n\n# Modeli ve tokenizer'\u0131 y\u00fckleme\n# device_map=\"auto\" ile model, mevcut GPU'lara otomatik olarak da\u011f\u0131t\u0131l\u0131r.\n# load_in_8bit=True veya load_in_4bit=True, daha az VRAM kullanmak i\u00e7in\n# modelin hassasiyetini d\u00fc\u015f\u00fcr\u00fcr. Bu, \u00f6zellikle b\u00fcy\u00fck modeller i\u00e7in kritiktir.\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n model_name,\n torch_dtype=torch.bfloat16, # veya torch.float16\n device_map=\"auto\",\n load_in_8bit=True # VRAM tasarrufu i\u00e7in 8-bit y\u00fckleme\n)\n\nprint(\"Model ba\u015far\u0131yla y\u00fcklendi ve kullan\u0131ma haz\u0131r!\")\nprint(f\"Modelin kulland\u0131\u011f\u0131 cihaz: {model.device}\")\n<\/code><\/pre>\n<p>\n Yukar\u0131daki kod blo\u011funda, <code class=\"language-\">model_name<\/code> de\u011fi\u015fkenini kullanmak istedi\u011finiz Nemotron 3 modelinin tam ad\u0131yla de\u011fi\u015ftirmeniz gerekmektedir. <code class=\"language-\">device_map=\"auto\"<\/code> parametresi, modelin GPU’nuzda veya birden fazla GPU’nuz varsa bunlar aras\u0131nda otomatik olarak da\u011f\u0131t\u0131lmas\u0131n\u0131 sa\u011flar. Bu, \u00f6zellikle modelin tek bir GPU’ya s\u0131\u011fmad\u0131\u011f\u0131 durumlarda \u00e7ok faydal\u0131d\u0131r. <code class=\"language-\">load_in_8bit=True<\/code> veya <code class=\"language-\">load_in_4bit=True<\/code> parametreleri ise, modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 daha d\u00fc\u015f\u00fck hassasiyetle (8-bit veya 4-bit) y\u00fckleyerek VRAM kullan\u0131m\u0131n\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r. Bu, daha b\u00fcy\u00fck modelleri daha az VRAM’e sahip GPU’larda \u00e7al\u0131\u015ft\u0131rman\u0131za olanak tan\u0131r, ancak k\u00fc\u00e7\u00fck bir performans d\u00fc\u015f\u00fc\u015f\u00fcne veya do\u011fruluk kayb\u0131na neden olabilir. Genellikle, bu d\u00fc\u015f\u00fc\u015f \u00e7o\u011fu uygulama i\u00e7in kabul edilebilir d\u00fczeydedir.\n<\/p>\n<p>\n <code class=\"language-\">torch_dtype=torch.bfloat16<\/code> veya <code class=\"language-\">torch.float16<\/code> ayar\u0131, modelin hangi veri tipiyle \u00e7al\u0131\u015faca\u011f\u0131n\u0131 belirler. <code class=\"language-\">bfloat16<\/code> (Brain Floating Point) veya <code class=\"language-\">float16<\/code> (Half Precision Floating Point) kullanmak, <code class=\"language-\">float32<\/code>‘ye k\u0131yasla daha az bellek t\u00fcketir ve modern GPU’larda daha h\u0131zl\u0131 hesaplamalar sa\u011flar. \u00c7o\u011fu yeni nesil GPU, bu hassasiyetleri donan\u0131msal olarak destekler. Modelin y\u00fcklenmesi, internet ba\u011flant\u0131n\u0131z\u0131n h\u0131z\u0131na ve modelin boyutuna ba\u011fl\u0131 olarak birka\u00e7 dakika s\u00fcrebilir. \u0130lk y\u00fcklemede model a\u011f\u0131rl\u0131klar\u0131 yerel diskinize indirilir, bu da sonraki y\u00fcklemelerin daha h\u0131zl\u0131 olmas\u0131n\u0131 sa\u011flar. Model ba\u015far\u0131yla y\u00fcklendikten sonra, <code class=\"language-\">model.device<\/code> \u00e7\u0131kt\u0131s\u0131 size modelin hangi cihazda (\u00f6rne\u011fin <code class=\"language-\">cuda:0<\/code>) \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 g\u00f6sterecektir. Bu a\u015famada, GPU’nuzun VRAM kullan\u0131m\u0131n\u0131 <code class=\"language-\">nvidia-smi<\/code> komutuyla kontrol edebilirsiniz. Modelin bellekte ne kadar yer kaplad\u0131\u011f\u0131n\u0131 g\u00f6rmek, optimizasyon ad\u0131mlar\u0131 i\u00e7in size fikir verecektir. Art\u0131k Nemotron 3 modeliniz GPU Droplet’inizde haz\u0131r ve metin \u00fcretimi gibi g\u00f6revler i\u00e7in kullan\u0131lmaya ba\u015flanabilir!\n<\/p>\n<h2>Modeli \u00c7al\u0131\u015ft\u0131rma ve \u00c7\u0131kar\u0131m Yapma: \u0130lk Metin \u00dcretiminiz<\/h2>\n<p>\n Nemotron 3 modelini GPU Droplet’inize ba\u015far\u0131yla y\u00fckledikten sonra, art\u0131k onunla etkile\u015fime ge\u00e7meye ve metin \u00fcretimi gibi \u00e7\u0131kar\u0131m (inference) g\u00f6revlerini ger\u00e7ekle\u015ftirmeye haz\u0131rs\u0131n\u0131z. Bu b\u00f6l\u00fcm, y\u00fckledi\u011finiz Nemotron 3 modelini kullanarak nas\u0131l metin olu\u015fturaca\u011f\u0131n\u0131z\u0131 ad\u0131m ad\u0131m g\u00f6sterecek ve \u00e7e\u015fitli parametrelerle \u00e7\u0131kt\u0131y\u0131 nas\u0131l kontrol edece\u011finizi a\u00e7\u0131klayacakt\u0131r. Modelden metin \u00fcretmek i\u00e7in temel olarak bir “prompt” (giri\u015f metni) sa\u011flaman\u0131z ve modelin bu prompt’u tamamlamas\u0131n\u0131 beklemeniz gerekir. Hugging Face Transformers k\u00fct\u00fcphanesi, bu s\u00fcreci olduk\u00e7a basitle\u015ftirir.\n<\/p>\n<pre><code class=\"language-python\">\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport torch\n\n# \u00d6nceki ad\u0131mdan y\u00fckledi\u011finiz modeli ve tokenizer'\u0131 tekrar y\u00fckleyin veya mevcut oturumunuzda kullan\u0131n\nmodel_name = \"nvidia\/Nemotron-3-8B-Instruct\" \n\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n model_name,\n torch_dtype=torch.bfloat16, \n device_map=\"auto\",\n load_in_8bit=True \n)\n\n# Metin \u00fcretimi i\u00e7in bir prompt haz\u0131rlay\u0131n\nprompt = \"Yapay zeka gelece\u011fi nas\u0131l \u015fekillendirecek?\"\n\n# Prompt'u token'lara d\u00f6n\u00fc\u015ft\u00fcr\u00fcn\ninput_ids = tokenizer.encode(prompt, return_tensors=\"pt\").to(model.device)\n\n# Metin \u00fcretimi i\u00e7in <code class=\"language-\">generate<\/code> metodunu kullan\u0131n\n# max_new_tokens: \u00dcretilecek maksimum yeni token say\u0131s\u0131\n# num_return_sequences: Ka\u00e7 farkl\u0131 \u00e7\u0131kt\u0131 \u00fcretilece\u011fi\n# do_sample: Rastgele \u00f6rnekleme yap\u0131l\u0131p yap\u0131lmayaca\u011f\u0131 (True ise daha yarat\u0131c\u0131, False ise daha deterministik)\n# temperature: Rastgelelik derecesi (daha y\u00fcksek de\u011ferler daha yarat\u0131c\u0131 ama tutars\u0131z olabilir)\n# top_k: En olas\u0131 K token aras\u0131ndan se\u00e7im yapma\n# top_p: K\u00fcm\u00fclatif olas\u0131l\u0131\u011f\u0131 P olan en olas\u0131 tokenlar aras\u0131ndan se\u00e7im yapma (\u00e7ekirdek \u00f6rnekleme)\noutput = model.generate(\n input_ids,\n max_new_tokens=200,\n num_return_sequences=1,\n do_sample=True,\n temperature=0.7,\n top_k=50,\n top_p=0.95,\n pad_token_id=tokenizer.eos_token_id # Modelin son token'\u0131n\u0131 belirtmek \u00f6nemlidir\n)\n\n# \u00dcretilen metni decode edin ve yazd\u0131r\u0131n\ngenerated_text = tokenizer.decode(output[0], skip_special_tokens=True)\nprint(\"--- \u00dcretilen Metin ---\")\nprint(generated_text)\nprint(\"----------------------\")\n<\/code><\/pre>\n<p>\n Yukar\u0131daki kod, temel bir metin \u00fcretim s\u00fcrecini g\u00f6stermektedir. <code class=\"language-\">model.generate<\/code> metodu, b\u00fcy\u00fck dil modellerinden metin \u00fcretmek i\u00e7in olduk\u00e7a g\u00fc\u00e7l\u00fc ve esnek bir ara\u00e7t\u0131r. \u00c7e\u015fitli parametreler sayesinde modelin davran\u0131\u015f\u0131n\u0131 ve \u00e7\u0131kt\u0131n\u0131n kalitesini kontrol edebilirsiniz:\n<\/p>\n<ul>\n<li><code class=\"language-\">max_new_tokens<\/code>: Modelin \u00fcretece\u011fi maksimum yeni token say\u0131s\u0131n\u0131 belirler. Bu, \u00e7\u0131kt\u0131n\u0131n uzunlu\u011funu kontrol etmek i\u00e7in \u00f6nemlidir.<\/li>\n<li><code class=\"language-\">num_return_sequences<\/code>: Ayn\u0131 prompt i\u00e7in ka\u00e7 farkl\u0131 metin \u00e7\u0131kt\u0131s\u0131 almak istedi\u011finizi belirtir. Bu, farkl\u0131 yarat\u0131c\u0131 alternatifler elde etmek i\u00e7in kullan\u0131\u015fl\u0131d\u0131r.<\/li>\n<li><code class=\"language-\">do_sample<\/code>: E\u011fer <code class=\"language-\">True<\/code> olarak ayarlan\u0131rsa, model rastgele \u00f6rnekleme tekniklerini kullanarak daha \u00e7e\u015fitli ve yarat\u0131c\u0131 \u00e7\u0131kt\u0131lar \u00fcretir. <code class=\"language-\">False<\/code> ise, model her zaman en olas\u0131 token’\u0131 se\u00e7erek daha deterministik \u00e7\u0131kt\u0131lar verir.<\/li>\n<li><code class=\"language-\">temperature<\/code>: <code class=\"language-\">do_sample=True<\/code> oldu\u011funda kullan\u0131l\u0131r. Daha y\u00fcksek s\u0131cakl\u0131k de\u011ferleri, modelin daha “cesur” se\u00e7imler yapmas\u0131na neden olarak daha yarat\u0131c\u0131 ama bazen tutars\u0131z \u00e7\u0131kt\u0131lar \u00fcretebilir. D\u00fc\u015f\u00fck de\u011ferler ise daha odakl\u0131 ve muhafazakar \u00e7\u0131kt\u0131lar verir.<\/li>\n<li><code class=\"language-\">top_k<\/code>: Modelin bir sonraki token’\u0131 se\u00e7erken sadece en olas\u0131 K token aras\u0131ndan se\u00e7im yapmas\u0131n\u0131 sa\u011flar. Bu, \u00e7\u0131kt\u0131n\u0131n kalitesini art\u0131rabilir ve anlams\u0131z token’lar\u0131n \u00fcretilmesini engeller.<\/li>\n<li><code class=\"language-\">top_p<\/code>: (\u00c7ekirdek \u00f6rnekleme olarak da bilinir) En olas\u0131 token’lar\u0131, k\u00fcm\u00fclatif olas\u0131l\u0131klar\u0131 <code class=\"language-\">top_p<\/code> de\u011ferine ula\u015fana kadar se\u00e7er. Bu, <code class=\"language-\">top_k<\/code>‘ye benzer \u015fekilde \u00e7\u0131kt\u0131n\u0131n kalitesini kontrol eder ancak daha dinamik bir token seti sa\u011flar.<\/li>\n<li><code class=\"language-\">pad_token_id<\/code>: Modelin “padding” token’\u0131n\u0131 belirtir. \u00d6zellikle toplu (batch) \u00e7\u0131kar\u0131m yaparken veya modelin sonland\u0131rma token’\u0131n\u0131 anlamas\u0131 i\u00e7in \u00f6nemlidir. Nemotron 3 genellikle <code class=\"language-\">eos_token_id<\/code> (end-of-sequence token) kullan\u0131r.<\/li>\n<\/ul>\n<p>\n Bu parametreleri deneyerek, Nemotron 3 modelinin farkl\u0131 senaryolara nas\u0131l tepki verdi\u011fini g\u00f6zlemleyebilir ve projeniz i\u00e7in en uygun \u00e7\u0131kt\u0131lar\u0131 elde edebilirsiniz. \u00d6rne\u011fin, bir hikaye olu\u015fturuyorsan\u0131z daha y\u00fcksek <code class=\"language-\">temperature<\/code> ve <code class=\"language-\">do_sample=True<\/code> kullanabilirken, bir kod a\u00e7\u0131klamas\u0131 istiyorsan\u0131z daha d\u00fc\u015f\u00fck <code class=\"language-\">temperature<\/code> ve <code class=\"language-\">top_k<\/code>\/<code class=\"language-\">top_p<\/code> de\u011ferleri tercih edebilirsiniz. Unutmay\u0131n, her model ve g\u00f6rev i\u00e7in en iyi parametre kombinasyonu farkl\u0131l\u0131k g\u00f6sterebilir, bu y\u00fczden deneme yapmak anahtard\u0131r.\n<\/p>\n<h2>Performans Optimizasyonu ve \u0130leri D\u00fczey \u0130pu\u00e7lar\u0131: Nemotron 3’\u00fc En Verimli \u015eekilde Kullanma<\/h2>\n<p>\n Nemotron 3 gibi b\u00fcy\u00fck dil modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, \u00f6nemli bir hesaplama ve bellek kayna\u011f\u0131 gerektirir. Bu nedenle, performans\u0131 optimize etmek ve kaynaklar\u0131 en verimli \u015fekilde kullanmak, hem maliyetleri d\u00fc\u015f\u00fcrmek hem de daha h\u0131zl\u0131 sonu\u00e7lar elde etmek i\u00e7in hayati \u00f6neme sahiptir. Bu b\u00f6l\u00fcmde, Nemotron 3 modellerini daha verimli \u00e7al\u0131\u015ft\u0131rmak i\u00e7in uygulayabilece\u011finiz ileri d\u00fczey teknikleri ve ipu\u00e7lar\u0131n\u0131 ele alaca\u011f\u0131z.\n<\/p>\n<h3>Quantization (Kuantizasyon) ile Bellek ve H\u0131z Kazanc\u0131 Nas\u0131l Sa\u011flan\u0131r?<\/h3>\n<p>\n Kuantizasyon, bir modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 ve aktivasyonlar\u0131n\u0131 daha d\u00fc\u015f\u00fck hassasiyetli veri tiplerine d\u00f6n\u00fc\u015ft\u00fcrme i\u015flemidir (\u00f6rne\u011fin, 32-bit kayan noktalardan 8-bit veya 4-bit tam say\u0131lara). Bu, modelin bellekte kaplad\u0131\u011f\u0131 alan\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r ve ayn\u0131 zamanda \u00e7\u0131kar\u0131m h\u0131z\u0131n\u0131 art\u0131rabilir, \u00e7\u00fcnk\u00fc d\u00fc\u015f\u00fck hassasiyetli hesaplamalar daha h\u0131zl\u0131d\u0131r. \u00d6zellikle <code class=\"language-\">bitsandbytes<\/code> k\u00fct\u00fcphanesi, bu i\u015flemi Hugging Face Transformers ile sorunsuz bir \u015fekilde entegre etmeyi sa\u011flar.\n<\/p>\n<pre><code class=\"language-python\">\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport torch\n\nmodel_name = \"nvidia\/Nemotron-3-8B-Instruct\" \n\n# Modelin 4-bit hassasiyetle y\u00fcklenmesi\n# Bu, VRAM kullan\u0131m\u0131n\u0131 %75'e kadar azaltabilir.\nmodel = AutoModelForCausalLM.from_pretrained(\n model_name,\n torch_dtype=torch.bfloat16,\n device_map=\"auto\",\n load_in_4bit=True # 4-bit kuantizasyon\n)\ntokenizer = AutoTokenizer.from_pretrained(model_name)\n\nprint(f\"Model 4-bit hassasiyetle y\u00fcklendi. Bellek kullan\u0131m\u0131 optimize edildi.\")\n<\/code><\/pre>\n<p>\n <code class=\"language-\">load_in_4bit=True<\/code> parametresi, modelin 4-bit kuantize edilmi\u015f versiyonunu y\u00fckleyerek VRAM kullan\u0131m\u0131n\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde d\u00fc\u015f\u00fcr\u00fcr. Bu, daha b\u00fcy\u00fck Nemotron 3 modellerini, daha az VRAM’e sahip GPU’larda bile \u00e7al\u0131\u015ft\u0131rman\u0131za olanak tan\u0131r. Kuantizasyonun k\u00fc\u00e7\u00fck bir do\u011fruluk kayb\u0131na neden olabilece\u011fini unutmay\u0131n, ancak \u00e7o\u011fu uygulama i\u00e7in bu kabul edilebilir d\u00fczeydedir.\n<\/p>\n<h3>Batch Inference (Toplu \u00c7\u0131kar\u0131m) Kullanarak \u0130\u015flem H\u0131z\u0131n\u0131 Art\u0131rma<\/h3>\n<p>\n Tek tek prompt’lar yerine birden fazla prompt’u ayn\u0131 anda i\u015flemek, GPU’nun paralel i\u015flem yeteneklerini daha iyi kullanmas\u0131n\u0131 sa\u011flar ve toplam \u00e7\u0131kar\u0131m s\u00fcresini azalt\u0131r. Buna “toplu \u00e7\u0131kar\u0131m” veya “batch inference” denir.\n<\/p>\n<pre><code class=\"language-python\">\n# Birden fazla prompt ile toplu \u00e7\u0131kar\u0131m \u00f6rne\u011fi\nprompts = [\n \"Yapay zeka gelece\u011fi nas\u0131l \u015fekillendirecek?\",\n \"Kuantum bilgisayarlar\u0131n potansiyeli nedir?\",\n \"\u0130klim de\u011fi\u015fikli\u011fi ile m\u00fccadelede yapay zekan\u0131n rol\u00fc.\"\n]\n\n# Prompt'lar\u0131 token'lara d\u00f6n\u00fc\u015ft\u00fcr\u00fcn ve pad edin\n# padding=True, truncation=True, return_tensors=\"pt\" ile t\u00fcm prompt'lar ayn\u0131 uzunlu\u011fa getirilir.\ninput_ids = tokenizer(prompts, return_tensors=\"pt\", padding=True, truncation=True).to(model.device)\n\noutput = model.generate(\n input_ids[\"input_ids\"],\n attention_mask=input_ids[\"attention_mask\"], # Dikkat maskesi de \u00f6nemlidir\n max_new_tokens=150,\n do_sample=True,\n temperature=0.8,\n top_p=0.9\n)\n\nfor i, out in enumerate(output):\n generated_text = tokenizer.decode(out, skip_special_tokens=True)\n print(f\"--- Prompt {i+1} ---\")\n print(generated_text)\n print(\"----------------------\")\n<\/code><\/pre>\n<p>\n Toplu \u00e7\u0131kar\u0131m yaparken, t\u00fcm giri\u015f prompt’lar\u0131n\u0131n ayn\u0131 uzunlukta olmas\u0131 i\u00e7in <code class=\"language-\">tokenizer<\/code>‘\u0131n <code class=\"language-\">padding=True<\/code> ve <code class=\"language-\">truncation=True<\/code> parametreleriyle kullan\u0131lmas\u0131 \u00f6nemlidir. Ayr\u0131ca, <code class=\"language-\">attention_mask<\/code>‘\u0131 da <code class=\"language-\">model.generate<\/code> metoduna iletmek, modelin padding token’lar\u0131n\u0131 g\u00f6z ard\u0131 etmesini sa\u011flar.\n<\/p>\n<h3>GPU VRAM Y\u00f6netimi ve Bellek \u0130pu\u00e7lar\u0131<\/h3>\n<p>\n B\u00fcy\u00fck modellerle \u00e7al\u0131\u015f\u0131rken VRAM y\u00f6netimi kritik \u00f6neme sahiptir.\n<\/p>\n<ul>\n<li><strong>Belle\u011fi Temizleme:<\/strong> Her \u00e7\u0131kar\u0131m i\u015fleminden sonra gereksiz belle\u011fi serbest b\u0131rakmak i\u00e7in <code class=\"language-\">torch.cuda.empty_cache()<\/code> kullan\u0131n.<\/li>\n<li><strong>Model Paralelle\u015ftirme:<\/strong> E\u011fer birden fazla GPU’nuz varsa, <code class=\"language-\">device_map=\"auto\"<\/code> veya <code class=\"language-\">model.parallelize()<\/code> gibi y\u00f6ntemlerle modeli GPU’lar aras\u0131nda da\u011f\u0131tarak daha b\u00fcy\u00fck modelleri \u00e7al\u0131\u015ft\u0131rabilirsiniz.<\/li>\n<li><strong>Gradient Checkpointing:<\/strong> \u0130nce ayar (fine-tuning) yaparken bellek kullan\u0131m\u0131n\u0131 azaltmak i\u00e7in <code class=\"language-\">gradient_checkpointing=True<\/code> parametresini kullanabilirsiniz. Bu, baz\u0131 hesaplamalar\u0131 yeniden yaparak bellek t\u00fcketimini azalt\u0131r, ancak e\u011fitim s\u00fcresini biraz uzatabilir.<\/li>\n<li><strong>Offloading:<\/strong> Baz\u0131 katmanlar\u0131 GPU’dan CPU’ya aktararak VRAM’i bo\u015faltma tekni\u011fidir. <code class=\"language-\">device_map=\"auto\"<\/code> bu i\u015flemi otomatik olarak yapabilir.<\/li>\n<\/ul>\n<h3>\u0130leri D\u00fczey \u00c7\u0131kar\u0131m Teknikleri<\/h3>\n<p>\n Daha h\u0131zl\u0131 \u00e7\u0131kar\u0131m i\u00e7in \u015funlar\u0131 d\u00fc\u015f\u00fcnebilirsiniz:\n<\/p>\n<ul>\n<li><strong>Flash Attention:<\/strong> \u00d6zellikle uzun dizilerle \u00e7al\u0131\u015f\u0131rken bellek ve h\u0131z a\u00e7\u0131s\u0131ndan \u00f6nemli kazan\u00e7lar sa\u011flayan optimize edilmi\u015f bir dikkat mekanizmas\u0131d\u0131r. Hugging Face <code class=\"language-\">transformers<\/code> k\u00fct\u00fcphanesi, uygun donan\u0131m ve yaz\u0131l\u0131m kurulumuyla Flash Attention 2’yi destekler.<\/li>\n<li><strong>\u00d6zel \u00c7\u0131kar\u0131m Motorlar\u0131:<\/strong> NVIDIA Triton Inference Server, ONNX Runtime veya TensorRT gibi \u00f6zel \u00e7\u0131kar\u0131m motorlar\u0131, modelleri \u00fcretim ortam\u0131nda \u00e7ok daha h\u0131zl\u0131 ve verimli bir \u015fekilde \u00e7al\u0131\u015ft\u0131rmak i\u00e7in optimize edilmi\u015f \u00e7\u00f6z\u00fcmler sunar. Bu ara\u00e7lar, modelinizi farkl\u0131 formatlara d\u00f6n\u00fc\u015ft\u00fcrerek donan\u0131ma \u00f6zel optimizasyonlar uygulayabilir.<\/li>\n<\/ul>\n<p>\n Bu optimizasyon tekniklerini uygulayarak, GPU Droplet’inizdeki Nemotron 3 modellerinden maksimum verimi alabilir, daha h\u0131zl\u0131 sonu\u00e7lar elde edebilir ve maliyetleri kontrol alt\u0131nda tutabilirsiniz. Her proje ve model i\u00e7in en uygun kombinasyonu bulmak i\u00e7in farkl\u0131 yakla\u015f\u0131mlar\u0131 denemek \u00f6nemlidir.\n<\/p>\n<h2>Vaka Analizi: Ger\u00e7ek D\u00fcnya Uygulamalar\u0131nda Nemotron 3 ve GPU Droplet<\/h2>\n<p>\n Nemotron 3 modelleri ve GPU Droplet’lerin birle\u015fimi, \u00e7e\u015fitli end\u00fcstrilerde ger\u00e7ek d\u00fcnya problemlerini \u00e7\u00f6zmek i\u00e7in g\u00fc\u00e7l\u00fc bir platform sunar. Bu b\u00f6l\u00fcmde, bu kombinasyonun nas\u0131l kullan\u0131labilece\u011fine dair iki farkl\u0131 vaka analizini inceleyece\u011fiz: bir m\u00fc\u015fteri hizmetleri botu ve i\u00e7erik olu\u015fturma asistan\u0131. Bu \u00f6rnekler, teorik bilgiyi pratik uygulamalara d\u00f6n\u00fc\u015ft\u00fcrmenize yard\u0131mc\u0131 olacakt\u0131r.\n<\/p>\n<h3>Vaka Analizi 1: Ak\u0131ll\u0131 M\u00fc\u015fteri Hizmetleri Botu<\/h3>\n<p>\n Bir e-ticaret \u015firketi, artan m\u00fc\u015fteri sorgular\u0131 nedeniyle mevcut m\u00fc\u015fteri hizmetleri ekibinin y\u00fck\u00fcn\u00fc hafifletmek ve 7\/24 destek sa\u011flamak istiyor. \u015eirket, hassas m\u00fc\u015fteri verilerini \u00fc\u00e7\u00fcnc\u00fc taraf API’lara g\u00f6ndermek istemedi\u011fi i\u00e7in kendi kontrol\u00fcnde, \u00f6zel bir yapay zeka \u00e7\u00f6z\u00fcm\u00fc aramaktad\u0131r.\n<\/p>\n<h4>Problem:<\/h4>\n<ul>\n<li>M\u00fc\u015fteri hizmetleri ekibinin i\u015f y\u00fck\u00fc y\u00fcksek.<\/li>\n<li>M\u00fc\u015fteri sorgular\u0131na h\u0131zl\u0131 ve tutarl\u0131 yan\u0131tlar verilemiyor.<\/li>\n<li>Hassas m\u00fc\u015fteri verilerinin g\u00fcvenli\u011fi \u00f6nemli.<\/li>\n<li>Mevcut \u00e7\u00f6z\u00fcmler ya \u00e7ok maliyetli ya da yeterince \u00f6zelle\u015ftirilebilir de\u011fil.<\/li>\n<\/ul>\n<h4>Nemotron 3 ve GPU Droplet \u00c7\u00f6z\u00fcm\u00fc:<\/h4>\n<p>\n \u015eirket, DigitalOcean’dan 24 GB VRAM’e sahip bir GPU Droplet kiralad\u0131 ve \u00fczerine Nemotron-3-8B-Instruct modelini kurdu. Model, \u015firket i\u00e7i SSS (S\u0131k\u00e7a Sorulan Sorular) belgeleri, \u00fcr\u00fcn k\u0131lavuzlar\u0131 ve ge\u00e7mi\u015f m\u00fc\u015fteri hizmetleri kay\u0131tlar\u0131 ile ince ayar (fine-tuning) yap\u0131ld\u0131. \u0130nce ayar s\u00fcreci, modelin \u015firketin \u00f6zel terminolojisini ve yan\u0131t tarz\u0131n\u0131 \u00f6\u011frenmesini sa\u011flad\u0131.\n<\/p>\n<pre><code class=\"language-python\">\n# \u0130nce ayarl\u0131 Nemotron 3 modelini y\u00fckleme (\u00f6rnek)\n# Bu k\u0131s\u0131m, ince ayar yap\u0131lm\u0131\u015f bir modelin y\u00fcklenmesini simgeler.\n# Ger\u00e7ek bir senaryoda, ince ayar sonras\u0131 model a\u011f\u0131rl\u0131klar\u0131 kaydedilir ve buradan y\u00fcklenir.\n# \u00d6rne\u011fin, \"my_company\/Nemotron-3-8B-Customer-Service\" gibi \u00f6zel bir model ad\u0131 olabilir.\n\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport torch\n\nmodel_name = \"my_company\/Nemotron-3-8B-Customer-Service\" # \u0130nce ayarl\u0131 model ad\u0131n\u0131z\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n model_name,\n torch_dtype=torch.bfloat16,\n device_map=\"auto\",\n load_in_8bit=True\n)\n\ndef get_customer_service_response(query):\n prompt = f\"M\u00fc\u015fteri: {query}\\nAsistan:\"\n input_ids = tokenizer.encode(prompt, return_tensors=\"pt\").to(model.device)\n output = model.generate(\n input_ids,\n max_new_tokens=100,\n do_sample=True,\n temperature=0.5,\n top_k=50,\n top_p=0.9\n )\n response = tokenizer.decode(output[0], skip_special_tokens=True)\n # Sadece asistan\u0131n yan\u0131t\u0131n\u0131 almak i\u00e7in prompt k\u0131sm\u0131n\u0131 \u00e7\u0131kar\u0131n\n return response.split(\"Asistan:\")[1].strip()\n\n# \u00d6rnek kullan\u0131m\nprint(get_customer_service_response(\"Sipari\u015fimin durumu nedir?\"))\nprint(get_customer_service_response(\"\u00dcr\u00fcn iade ko\u015fullar\u0131 nelerdir?\"))\n<\/code><\/pre>\n<h4>Sonu\u00e7lar:<\/h4>\n<ul>\n<li>M\u00fc\u015fteri sorgular\u0131n\u0131n %70’ine yak\u0131n bir k\u0131sm\u0131 bot taraf\u0131ndan an\u0131nda ve do\u011fru bir \u015fekilde yan\u0131tland\u0131.<\/li>\n<li>M\u00fc\u015fteri hizmetleri ekibi, daha karma\u015f\u0131k ve \u00f6zel durumlarla ilgilenmeye odaklanabildi.<\/li>\n<li>Yan\u0131t s\u00fcreleri \u00f6nemli \u00f6l\u00e7\u00fcde azald\u0131 ve m\u00fc\u015fteri memnuniyeti artt\u0131.<\/li>\n<li>\u015eirket, verilerinin kendi sunucular\u0131nda kalmas\u0131ndan dolay\u0131 g\u00fcvenlik endi\u015fesi ya\u015famad\u0131.<\/li>\n<\/ul>\n<h3>Vaka Analizi 2: Pazarlama \u0130\u00e7erik Olu\u015fturma Asistan\u0131<\/h3>\n<p>\n Bir dijital pazarlama ajans\u0131, m\u00fc\u015fterileri i\u00e7in s\u00fcrekli olarak blog yaz\u0131lar\u0131, sosyal medya g\u00f6nderileri ve reklam metinleri \u00fcretmek zorundad\u0131r. Bu s\u00fcre\u00e7 zaman al\u0131c\u0131 ve maliyetlidir. Ajans, yarat\u0131c\u0131 engelleri a\u015fmak ve i\u00e7erik \u00fcretimini h\u0131zland\u0131rmak i\u00e7in bir yapay zeka asistan\u0131 aray\u0131\u015f\u0131ndad\u0131r.\n<\/p>\n<h4>Problem:<\/h4>\n<ul>\n<li>Yarat\u0131c\u0131 i\u00e7erik \u00fcretimi zaman al\u0131c\u0131 ve kaynak yo\u011fun.<\/li>\n<li>Farkl\u0131 m\u00fc\u015fteriler i\u00e7in ton ve stil tutarl\u0131l\u0131\u011f\u0131n\u0131 sa\u011flamak zor.<\/li>\n<li>Yazar blo\u011fu gibi yarat\u0131c\u0131 engellerle s\u0131k\u00e7a kar\u015f\u0131la\u015f\u0131l\u0131yor.<\/li>\n<li>Pazarlama ekibinin daha stratejik g\u00f6revlere odaklanmas\u0131 gerekiyor.<\/li>\n<\/ul>\n<h4>Nemotron 3 ve GPU Droplet \u00c7\u00f6z\u00fcm\u00fc:<\/h4>\n<p>\n Ajans, daha b\u00fcy\u00fck bir Nemotron-3-22B-Base modelini bar\u0131nd\u0131rabilecek daha g\u00fc\u00e7l\u00fc bir GPU Droplet (\u00f6rne\u011fin A100 GPU’lu) se\u00e7ti. Model, ajans\u0131n ge\u00e7mi\u015fteki ba\u015far\u0131l\u0131 pazarlama kampanyalar\u0131, blog yaz\u0131lar\u0131 ve marka k\u0131lavuzlar\u0131 ile ince ayarland\u0131. Bu sayede model, ajans\u0131n belirli m\u00fc\u015fteri profillerine ve marka sesine uygun i\u00e7erikler \u00fcretebildi.\n<\/p>\n<pre><code class=\"language-python\">\n# Pazarlama i\u00e7erik asistan\u0131 i\u00e7in Nemotron 3 kullan\u0131m\u0131 (\u00f6rnek)\n\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport torch\n\nmodel_name = \"nvidia\/Nemotron-3-22B-Base\" # Daha b\u00fcy\u00fck bir model se\u00e7imi\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n model_name,\n torch_dtype=torch.bfloat16,\n device_map=\"auto\" # 22B model tek GPU'ya s\u0131\u011fmayabilir, otomatik da\u011f\u0131t\u0131m \u00f6nemli\n)\n\ndef generate_marketing_content(topic, content_type=\"blog post\", tone=\"informative\"):\n prompt = f\"Bir {tone} tonda, {topic} hakk\u0131nda bir {content_type} tasla\u011f\u0131 olu\u015ftur.\"\n input_ids = tokenizer.encode(prompt, return_tensors=\"pt\").to(model.device)\n output = model.generate(\n input_ids,\n max_new_tokens=300,\n do_sample=True,\n temperature=0.9, # Daha yarat\u0131c\u0131 \u00e7\u0131kt\u0131lar i\u00e7in y\u00fcksek s\u0131cakl\u0131k\n top_k=50,\n top_p=0.95,\n num_return_sequences=2 # \u0130ki farkl\u0131 taslak \u00fcret\n )\n generated_contents = [tokenizer.decode(out, skip_special_tokens=True) for out in output]\n return [content.split(prompt)[1].strip() for content in generated_contents]\n\n# \u00d6rnek kullan\u0131m\ntopics = [\"S\u00fcrd\u00fcr\u00fclebilir Moda Trendleri\", \"Yapay Zeka Destekli Pazarlama\"]\nfor topic in topics:\n contents = generate_marketing_content(topic, content_type=\"sosyal medya g\u00f6nderisi\", tone=\"ilham verici\")\n print(f\"\\n--- {topic} i\u00e7in Sosyal Medya G\u00f6nderisi Taslaklar\u0131 ---\")\n for i, content in enumerate(contents):\n print(f\"Taslak {i+1}:\\n{content}\\n\")\n<\/code><\/pre>\n<h4>Sonu\u00e7lar:<\/h4>\n<ul>\n<li>\u0130\u00e7erik \u00fcretim s\u00fcresi %40 oran\u0131nda azald\u0131.<\/li>\n<li>Farkl\u0131 m\u00fc\u015fteriler i\u00e7in tutarl\u0131 marka sesi ve tonu ile i\u00e7erikler \u00fcretilebildi.<\/li>\n<li>Pazarlama ekibi, i\u00e7erik olu\u015fturma yerine strateji geli\u015ftirme ve kampanya y\u00f6netimine daha fazla zaman ay\u0131rabildi.<\/li>\n<li>Yarat\u0131c\u0131 engeller a\u015f\u0131larak daha \u00e7e\u015fitli ve yenilik\u00e7i i\u00e7erik fikirleri ortaya \u00e7\u0131kt\u0131.<\/li>\n<\/ul>\n<p>\n Bu vaka analizleri, Nemotron 3 modellerinin GPU Droplet’ler \u00fczerinde nas\u0131l kullan\u0131labilece\u011fine dair sadece birka\u00e7 \u00f6rnektir. Do\u011fru model se\u00e7imi, ince ayar ve optimizasyon stratejileriyle, bu g\u00fc\u00e7l\u00fc kombinasyonun potansiyeli neredeyse s\u0131n\u0131rs\u0131zd\u0131r.\n<\/p>\n<h2>S\u0131k\u00e7a Sorulan Sorular (SSS)<\/h2>\n<h3>Nemotron 3 modelleri hangi dilleri destekliyor?<\/h3>\n<p>\n Nemotron 3 modelleri, \u00f6zellikle \u0130ngilizce \u00fczerinde yo\u011funla\u015fm\u0131\u015f olsa da, \u00e7ok dilli veri k\u00fcmeleri \u00fczerinde de e\u011fitilmi\u015ftir ve bir\u00e7ok farkl\u0131 dilde metin anlama ve \u00fcretme yetene\u011fine sahiptir. Ancak, \u0130ngilizce d\u0131\u015f\u0131ndaki dillerdeki performans\u0131, modelin boyutuna ve spesifik e\u011fitim verilerine ba\u011fl\u0131 olarak de\u011fi\u015febilir. En iyi performans i\u00e7in hedef dilinizde ince ayar yap\u0131lmas\u0131 \u00f6nerilir.\n<\/p>\n<h3>GPU Droplet’imin VRAM’i Nemotron 3 modelini \u00e7al\u0131\u015ft\u0131rmak i\u00e7in yeterli mi?<\/h3>\n<p>\n Bu, kullanmak istedi\u011finiz Nemotron 3 modelinin boyutuna ve uygulayaca\u011f\u0131n\u0131z optimizasyonlara (kuantizasyon gibi) ba\u011fl\u0131d\u0131r. \u00d6rne\u011fin, 8 milyar parametreli bir model i\u00e7in en az 16 GB VRAM \u00f6nerilirken, 22 milyar parametreli modeller 4-bit kuantizasyon ile bile 24 GB veya daha fazla VRAM gerektirebilir. <code class=\"language-\">nvidia-smi<\/code> komutuyla Droplet’inizin VRAM’ini kontrol edebilir ve modelinizi <code class=\"language-\">load_in_8bit=True<\/code> veya <code class=\"language-\">load_in_4bit=True<\/code> gibi parametrelerle y\u00fckleyerek bellek kullan\u0131m\u0131n\u0131 optimize edebilirsiniz.\n<\/p>\n<h3>Nemotron 3 modellerini ince ayar (fine-tuning) yapabilir miyim?<\/h3>\n<p>\n Evet, Nemotron 3 gibi a\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 modeller, kendi \u00f6zel veri setlerinizle ince ayar yapmaya uygundur. Bu, modelin belirli bir g\u00f6rev, sekt\u00f6r veya dil i\u00e7in performans\u0131n\u0131 ve tutarl\u0131l\u0131\u011f\u0131n\u0131 art\u0131rman\u0131za olanak tan\u0131r. \u0130nce ayar i\u00e7in genellikle daha fazla GPU kayna\u011f\u0131 ve \u00f6zel e\u011fitim d\u00f6ng\u00fcleri gereklidir, ancak bu, modelin uygulaman\u0131za \u00f6zel hale getirilmesi i\u00e7in g\u00fc\u00e7l\u00fc bir y\u00f6ntemdir. Hugging Face <code class=\"language-\">transformers<\/code> k\u00fct\u00fcphanesi, ince ayar s\u00fcre\u00e7lerini kolayla\u015ft\u0131ran ara\u00e7lar sunar.\n<\/p>\n<h3>GPU Droplet kullanman\u0131n maliyeti ne kadar?<\/h3>\n<p>\n GPU Droplet’lerin maliyeti, se\u00e7ti\u011finiz bulut sa\u011flay\u0131c\u0131s\u0131na (DigitalOcean, AWS, Google Cloud vb.), GPU t\u00fcr\u00fcne (\u00f6rn. A100, RTX 4000), VRAM miktar\u0131na, CPU \u00e7ekirdek say\u0131s\u0131na ve RAM’e ba\u011fl\u0131 olarak b\u00fcy\u00fck \u00f6l\u00e7\u00fcde de\u011fi\u015fir. Genellikle saatlik veya ayl\u0131k olarak faturaland\u0131r\u0131l\u0131rlar. Daha k\u00fc\u00e7\u00fck GPU’lu Droplet’ler saatlik birka\u00e7 dolardan ba\u015flarken, y\u00fcksek performansl\u0131 A100 gibi GPU’lar saatlik 10-20 dolar veya daha fazla olabilir. Maliyetleri optimize etmek i\u00e7in sadece ihtiyac\u0131n\u0131z oldu\u011funda Droplet’i \u00e7al\u0131\u015ft\u0131rmak ve i\u015finiz bitti\u011finde kapatmak \u00f6nemlidir.\n<\/p>\n<h3>Nemotron 3 modelleri i\u00e7in mobil uyumlu HTML \u00e7\u0131kt\u0131s\u0131 nas\u0131l sa\u011flan\u0131r?<\/h3>\n<p>\n Nemotron 3 modelleri do\u011frudan HTML \u00e7\u0131kt\u0131s\u0131 \u00fcretmez; metin tabanl\u0131 \u00e7\u0131kt\u0131lar sa\u011flar. Ancak, bu metin \u00e7\u0131kt\u0131lar\u0131n\u0131 mobil uyumlu bir web aray\u00fcz\u00fcnde g\u00f6stermek i\u00e7in modern web geli\u015ftirme tekniklerini kullanabilirsiniz. \u00d6rne\u011fin, yan\u0131tlar\u0131 bir web uygulamas\u0131na entegre ederken, CSS <code class=\"language-\">media query<\/code>‘ler kullanarak farkl\u0131 ekran boyutlar\u0131na uyum sa\u011flayan duyarl\u0131 (responsive) tasar\u0131mlar olu\u015fturabilirsiniz.\n<\/p>\n<pre><code class=\"language-css\">\n\/* \u00d6rnek bir CSS media query *\/\n@media (max-width: 768px) {\n body {\n font-size: 14px;\n }\n .container {\n padding: 10px;\n }\n .response-box {\n width: 100%;\n margin-bottom: 15px;\n }\n}\n<\/code><\/pre>\n<p>\n Bu \u00f6rnekte, ekran geni\u015fli\u011fi 768 pikselin alt\u0131na d\u00fc\u015ft\u00fc\u011f\u00fcnde font boyutlar\u0131, dolgu ve yan\u0131t kutular\u0131n\u0131n geni\u015fli\u011fi otomatik olarak mobil cihazlara uygun hale getirilir. Modelin \u00fcretti\u011fi metin i\u00e7eri\u011fini bir web sayfas\u0131na entegre ederken, bu t\u00fcr CSS kurallar\u0131yla birlikte <code class=\"language-\"><meta name=\"viewport\" content=\"width=device-width, initial-scale=1.0\"><\/code> etiketini HTML’in <code class=\"language-\"><head><\/code> b\u00f6l\u00fcm\u00fcne eklemek, taray\u0131c\u0131ya i\u00e7eri\u011fi cihaz\u0131n geni\u015fli\u011fine g\u00f6re \u00f6l\u00e7eklemesini s\u00f6yler.\n<\/p>\n<h2>Sonu\u00e7: Yapay Zekan\u0131n Gelece\u011fine Yolculuk<\/h2>\n<p>\n A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rma ser\u00fcvenimiz, yapay zeka teknolojilerinin ne kadar eri\u015filebilir ve g\u00fc\u00e7l\u00fc hale geldi\u011fini a\u00e7\u0131k\u00e7a ortaya koymaktad\u0131r. Bu rehber boyunca, temel kavramlardan ba\u015flayarak bir GPU Droplet’in nas\u0131l se\u00e7ilece\u011fi ve kurulaca\u011f\u0131, gerekli yaz\u0131l\u0131mlar\u0131n nas\u0131l y\u00fcklenece\u011fi, Nemotron 3 modelinin nas\u0131l indirilece\u011fi ve \u00e7al\u0131\u015ft\u0131r\u0131laca\u011f\u0131 ad\u0131mlar\u0131n\u0131 detayl\u0131 bir \u015fekilde inceledik. Ayr\u0131ca, performans optimizasyonu i\u00e7in kuantizasyon ve toplu \u00e7\u0131kar\u0131m gibi ileri d\u00fczey tekniklere de\u011findik ve ger\u00e7ek d\u00fcnya senaryolar\u0131yla vaka analizleri sunduk. Bu bilgilerle donanm\u0131\u015f olarak, kendi yapay zeka projelerinizi hayata ge\u00e7irmek i\u00e7in sa\u011flam bir temele sahipsiniz.\n<\/p>\n<p>\n Nemotron 3 gibi a\u00e7\u0131k kaynakl\u0131 modellerin sundu\u011fu esneklik ve GPU Droplet’lerin sa\u011flad\u0131\u011f\u0131 \u00f6l\u00e7eklenebilir, maliyet etkin hesaplama g\u00fcc\u00fc, geli\u015ftiricilere ve i\u015fletmelere daha \u00f6nce m\u00fcmk\u00fcn olmayan f\u0131rsatlar sunmaktad\u0131r. Kendi kontrol\u00fcn\u00fczdeki bir ortamda bu modelleri \u00e7al\u0131\u015ft\u0131rarak, veri gizlili\u011fini korurken ayn\u0131 zamanda yenilik\u00e7i uygulamalar geli\u015ftirebilirsiniz. Metin \u00fcretimi, sohbet botlar\u0131, i\u00e7erik otomasyonu, veri analizi ve daha fazlas\u0131 i\u00e7in Nemotron 3’\u00fcn potansiyeli s\u0131n\u0131rs\u0131zd\u0131r. Yapay zeka d\u00fcnyas\u0131 s\u00fcrekli evrim ge\u00e7irse de, bu t\u00fcr g\u00fc\u00e7l\u00fc ara\u00e7lar\u0131 etkin bir \u015fekilde kullanabilme yetene\u011fi, gelecekteki teknolojik ilerlemelerin anahtar\u0131 olacakt\u0131r. Unutmay\u0131n, en iyi \u00f6\u011frenme yolu deneme yan\u0131lmad\u0131r. Farkl\u0131 parametrelerle oynay\u0131n, farkl\u0131 modelleri deneyin ve kendi benzersiz \u00e7\u00f6z\u00fcmlerinizi ke\u015ffedin. Yapay zekan\u0131n gelece\u011fine olan yolculu\u011funuzda ba\u015far\u0131lar dileriz!\n<\/p>\n<p><\/body><\/p>\n","protected":false},"excerpt":{"rendered":"GPU Droplet’te Nemotron 3 Modelleri Nas\u0131l \u00c7al\u0131\u015ft\u0131r\u0131l\u0131r? A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay…","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1],"tags":[],"class_list":{"0":"post-37506","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\n<title>Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?<\/title>\n<meta name=\"description\" content=\"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\" \/>\n<meta property=\"og:locale\" content=\"tr_TR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?\" \/>\n<meta property=\"og:description\" content=\"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\" \/>\n<meta property=\"og:site_name\" content=\"Kodlar\u0131n Gizemli D\u00fcnyas\u0131\" \/>\n<meta property=\"article:published_time\" content=\"2026-01-06T21:21:06+00:00\" \/>\n<meta name=\"author\" content=\"Fatih Soysal\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Yazan:\" \/>\n\t<meta name=\"twitter:data1\" content=\"Fatih Soysal\" \/>\n\t<meta name=\"twitter:label2\" content=\"Tahmini okuma s\u00fcresi\" \/>\n\t<meta name=\"twitter:data2\" content=\"30 dakika\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\"},\"author\":{\"name\":\"Fatih Soysal\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"headline\":\"Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?\",\"datePublished\":\"2026-01-06T21:21:06+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\"},\"wordCount\":4841,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#respond\"]}],\"copyrightYear\":\"2026\",\"copyrightHolder\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#organization\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\",\"name\":\"Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\"},\"datePublished\":\"2026-01-06T21:21:06+00:00\",\"description\":\"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.\",\"breadcrumb\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#breadcrumb\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Anasayfa\",\"item\":\"https:\/\/fatihsoysal.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/\",\"name\":\"Fatihsoysal.com\",\"description\":\"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim\",\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"tr\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\",\"name\":\"Fatih Soysal\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"tr\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"contentUrl\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"width\":512,\"height\":512,\"caption\":\"Fatih Soysal\"},\"logo\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\"},\"description\":\"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/\"}]}<\/script>\n","yoast_head_json":{"title":"Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?","description":"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/","og_locale":"tr_TR","og_type":"article","og_title":"Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?","og_description":"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.","og_url":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/","og_site_name":"Kodlar\u0131n Gizemli D\u00fcnyas\u0131","article_published_time":"2026-01-06T21:21:06+00:00","author":"Fatih Soysal","twitter_card":"summary_large_image","twitter_misc":{"Yazan:":"Fatih Soysal","Tahmini okuma s\u00fcresi":"30 dakika"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#article","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/"},"author":{"name":"Fatih Soysal","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"headline":"Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?","datePublished":"2026-01-06T21:21:06+00:00","mainEntityOfPage":{"@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/"},"wordCount":4841,"commentCount":0,"publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"inLanguage":"tr","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#respond"]}],"copyrightYear":"2026","copyrightHolder":{"@id":"https:\/\/fatihsoysal.com\/blog\/#organization"}},{"@type":"WebPage","@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/","url":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/","name":"Nemotron 3 Modelleri ve GPU Droplet'lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/#website"},"datePublished":"2026-01-06T21:21:06+00:00","description":"A\u00e7\u0131k a\u011f\u0131rl\u0131kl\u0131 Nemotron 3 modellerini bir GPU Droplet \u00fczerinde \u00e7al\u0131\u015ft\u0131rmak, yapay zeka projeleriniz i\u00e7in g\u00fc\u00e7l\u00fc ve esnek \u00e7\u00f6z\u00fcmler sunar. Bu kapsaml\u0131 rehberde, bu s\u00fcre\u00e7te ad\u0131m ad\u0131m ilerleyerek, modelleri da\u011f\u0131tma ve optimize etme y\u00f6ntemlerini ke\u015ffedeceksiniz.","breadcrumb":{"@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#breadcrumb"},"inLanguage":"tr","potentialAction":[{"@type":"ReadAction","target":["https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Anasayfa","item":"https:\/\/fatihsoysal.com\/blog\/"},{"@type":"ListItem","position":2,"name":"Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?"}]},{"@type":"WebSite","@id":"https:\/\/fatihsoysal.com\/blog\/#website","url":"https:\/\/fatihsoysal.com\/blog\/","name":"Fatihsoysal.com","description":"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim","publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"tr"},{"@type":["Person","Organization"],"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1","name":"Fatih Soysal","image":{"@type":"ImageObject","inLanguage":"tr","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/","url":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","contentUrl":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","width":512,"height":512,"caption":"Fatih Soysal"},"logo":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/"},"description":"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.","url":"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/"}]}},"yoast_meta":{"yoast_wpseo_title":"","yoast_wpseo_metadesc":"","yoast_wpseo_canonical":""},"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/37506","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/comments?post=37506"}],"version-history":[{"count":0,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/37506\/revisions"}],"wp:attachment":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/media?parent=37506"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/categories?post=37506"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/tags?post=37506"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}

{"id":37506,"date":"2026-01-07T00:21:06","date_gmt":"2026-01-06T21:21:06","guid":{"rendered":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/"},"modified":"2026-01-07T00:21:06","modified_gmt":"2026-01-06T21:21:06","slug":"nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli","status":"publish","type":"post","link":"https:\/\/fatihsoysal.com\/blog\/nemotron-3-modelleri-ve-gpu-dropletlerin-gucu-neden-bu-kombinasyon-onemli\/","title":{"rendered":"Nemotron 3 Modelleri ve GPU Droplet’lerin G\u00fcc\u00fc: Neden Bu Kombinasyon \u00d6nemli?"},"content":{"rendered":"