{"id":31185,"date":"2025-10-07T00:41:10","date_gmt":"2025-10-06T21:41:10","guid":{"rendered":"https:\/\/fatihsoysal.com\/blog\/?p=31185"},"modified":"2025-10-07T00:41:10","modified_gmt":"2025-10-06T21:41:10","slug":"deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber","status":"publish","type":"post","link":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/","title":{"rendered":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber"},"content":{"rendered":"<p><body><\/p>\n<h2>DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber<\/h2>\n<h3>Giri\u015f: B\u00fcy\u00fck Dil Modellerinin Da\u011f\u0131t\u0131m\u0131 ve Optimizasyonu<\/h3>\n<p>B\u00fcy\u00fck Dil Modelleri (LLM&#8217;ler), do\u011fal dil i\u015fleme alan\u0131nda devrim yaratarak metin \u00fcretimi, \u00e7eviri, \u00f6zetleme ve karma\u015f\u0131k sorular\u0131 yan\u0131tlama gibi bir\u00e7ok g\u00f6revi ba\u015far\u0131yla yerine getirmektedir. Ancak, bu modellerin \u00fcretim ortamlar\u0131nda da\u011f\u0131t\u0131m\u0131 ve y\u00f6netimi, y\u00fcksek hesaplama kaynaklar\u0131 gereksinimleri ve \u00e7\u0131kar\u0131m (inference) s\u00fcrecindeki gecikme s\u00fcreleri nedeniyle \u00f6nemli zorluklar bar\u0131nd\u0131r\u0131r. \u00d6zellikle, binlerce veya milyonlarca token \u00fcreten uzun ba\u011flaml\u0131 sorgular, GPU belle\u011fi ve hesaplama g\u00fcc\u00fc \u00fczerinde ciddi bir y\u00fck olu\u015fturur. Bu makale, DeepScaleR-1.5B gibi optimize edilmi\u015f, daha k\u00fc\u00e7\u00fck \u00f6l\u00e7ekli bir LLM&#8217;i, y\u00fcksek performansl\u0131 vLLM \u00e7\u0131kar\u0131m \u00e7er\u00e7evesi kullanarak DigitalOcean bulut platformunda nas\u0131l da\u011f\u0131taca\u011f\u0131m\u0131z\u0131 ve \u00e7al\u0131\u015ft\u0131raca\u011f\u0131m\u0131z\u0131 detayl\u0131 bir \u015fekilde ele almaktad\u0131r. Amac\u0131m\u0131z, maliyet etkinli\u011fi ve performans\u0131 dengeleyerek, \u00fcretim ortam\u0131na haz\u0131r, \u00f6l\u00e7eklenebilir bir LLM hizmeti sunmakt\u0131r.<\/p>\n<h3>B\u00fcy\u00fck Dil Modelleri (LLM&#8217;ler) ve DeepScaleR-1.5B&#8217;nin Yeri<\/h3>\n<p>B\u00fcy\u00fck Dil Modelleri, trilyonlarca parametreye ula\u015fabilen ve milyarlarca token \u00fczerinde e\u011fitilmi\u015f derin \u00f6\u011frenme modelleridir. Transformer mimarisine dayanan bu modeller, dikkat mekanizmas\u0131 sayesinde uzun mesafeli ba\u011f\u0131ml\u0131l\u0131klar\u0131 yakalayarak insan benzeri metinler \u00fcretebilirler. GPT-3, LLaMA, Falcon gibi modeller bu kategorinin \u00f6nde gelen \u00f6rnekleridir. Ancak, bu modellerin devasa boyutlar\u0131, \u00e7\u0131kar\u0131m s\u0131ras\u0131nda \u00f6nemli GPU belle\u011fi ve hesaplama g\u00fcc\u00fc gerektirir. Bu durum, \u00f6zellikle k\u0131s\u0131tl\u0131 b\u00fct\u00e7eye sahip geli\u015ftiriciler ve k\u00fc\u00e7\u00fck\/orta \u00f6l\u00e7ekli i\u015fletmeler i\u00e7in bir engel te\u015fkil edebilir.<\/p>\n<p>DeepScaleR-1.5B, bu zorluklar\u0131 a\u015fmak amac\u0131yla geli\u015ftirilmi\u015f, 1.5 milyar parametreye sahip, g\u00f6receli olarak daha k\u00fc\u00e7\u00fck ve optimize edilmi\u015f bir dil modelidir. Genellikle belirli g\u00f6revler veya diller i\u00e7in ince ayar yap\u0131lm\u0131\u015f (fine-tuned) versiyonlar\u0131 bulunan bu t\u00fcr modeller, daha d\u00fc\u015f\u00fck maliyetli donan\u0131mlarda bile kabul edilebilir performans sunarak geni\u015f bir kullan\u0131m alan\u0131 bulur. DeepScaleR-1.5B&#8217;nin temel avantajlar\u0131 \u015funlard\u0131r:<\/p>\n<ul>\n<li><strong>Daha D\u00fc\u015f\u00fck Kaynak T\u00fcketimi:<\/strong> Daha az parametre say\u0131s\u0131, daha az GPU belle\u011fi ve daha d\u00fc\u015f\u00fck hesaplama g\u00fcc\u00fc gereksinimi anlam\u0131na gelir.<\/li>\n<li><strong>H\u0131zl\u0131 \u00c7\u0131kar\u0131m:<\/strong> Daha k\u00fc\u00e7\u00fck boyut, ayn\u0131 donan\u0131mda daha y\u00fcksek token \u00fcretim h\u0131z\u0131 ve daha d\u00fc\u015f\u00fck gecikme s\u00fcresi sa\u011flar.<\/li>\n<li><strong>Maliyet Etkinli\u011fi:<\/strong> Daha uygun fiyatl\u0131 GPU&#8217;lar \u00fczerinde \u00e7al\u0131\u015fabilmesi, bulut hizmet maliyetlerini \u00f6nemli \u00f6l\u00e7\u00fcde d\u00fc\u015f\u00fcr\u00fcr.<\/li>\n<li><strong>\u00d6zelle\u015ftirilebilirlik:<\/strong> Belirli bir alana veya g\u00f6reve y\u00f6nelik olarak ince ayar yap\u0131lmas\u0131 daha kolay ve h\u0131zl\u0131d\u0131r.<\/li>\n<\/ul>\n<p>DeepScaleR-1.5B&#8217;yi se\u00e7memizin nedeni, y\u00fcksek kaliteli \u00e7\u0131kt\u0131lar sunarken ayn\u0131 zamanda DigitalOcean gibi maliyet-performans dengesi iyi olan bir platformda ekonomik bir \u015fekilde \u00e7al\u0131\u015ft\u0131r\u0131labilmesidir.<\/p>\n<h3>vLLM: Y\u00fcksek Performansl\u0131 LLM \u00c7\u0131kar\u0131m \u00c7er\u00e7evesi<\/h3>\n<p>LLM&#8217;lerin verimli bir \u015fekilde sunulmas\u0131, sadece modelin boyutuyla ilgili de\u011fildir; ayn\u0131 zamanda \u00e7\u0131kar\u0131m \u00e7er\u00e7evesinin optimizasyon yetenekleriyle de yak\u0131ndan ili\u015fkilidir. Geleneksel \u00e7\u0131kar\u0131m y\u00f6ntemleri, \u00f6zellikle birden fazla e\u015fzamanl\u0131 istek geldi\u011finde veya uzun metinler \u00fcretildi\u011finde performans darbo\u011fazlar\u0131 ya\u015fayabilir. \u0130\u015fte bu noktada vLLM devreye girer.<\/p>\n<p>vLLM, LLM \u00e7\u0131kar\u0131m\u0131n\u0131 h\u0131zland\u0131rmak ve optimize etmek i\u00e7in tasarlanm\u0131\u015f a\u00e7\u0131k kaynakl\u0131 bir k\u00fct\u00fcphanedir. Temel amac\u0131, GPU kullan\u0131m\u0131n\u0131 maksimize ederek token \u00fcretim h\u0131z\u0131n\u0131 (throughput) art\u0131rmak ve gecikme s\u00fcresini (latency) azaltmakt\u0131r. vLLM&#8217;in sundu\u011fu temel optimizasyonlar \u015funlard\u0131r:<\/p>\n<ul>\n<li><strong>PagedAttention:<\/strong> Bu, vLLM&#8217;in en \u00f6nemli yeniliklerinden biridir. Transformer modellerinde attention mekanizmas\u0131 i\u00e7in gerekli olan key ve value (KV) \u00f6nbelleklerini (cache) etkin bir \u015fekilde y\u00f6netir. Geleneksel yakla\u015f\u0131mlar, her sorgu i\u00e7in ayr\u0131lan KV \u00f6nbelle\u011finin tamam\u0131n\u0131 kullanmasa bile bellekte tutar. PagedAttention ise sanal bellek sistemlerindeki sayfalama (paging) mekanizmas\u0131na benzer \u015fekilde, KV \u00f6nbelle\u011fini k\u00fc\u00e7\u00fck bloklara ay\u0131r\u0131r ve yaln\u0131zca ger\u00e7ekten kullan\u0131lan bloklar\u0131 bellekte tutar. Bu sayede, GPU belle\u011fi \u00e7ok daha verimli kullan\u0131l\u0131r ve ayn\u0131 anda daha fazla e\u015fzamanl\u0131 istek i\u015flenebilir.<\/li>\n<li><strong>S\u00fcrekli Gruplama (Continuous Batching):<\/strong> Geleneksel gruplama y\u00f6ntemlerinde, bir grup tamamlanana kadar yeni istekler bekletilir. vLLM, s\u00fcrekli gruplama sayesinde, bir grubun i\u00e7indeki istekler tamamlan\u0131rken bile yeni istekleri dinamik olarak gruba ekleyebilir. Bu, GPU&#8217;nun bo\u015fta kalma s\u00fcresini azalt\u0131r ve genel verimi art\u0131r\u0131r.<\/li>\n<li><strong>\u00d6zelle\u015ftirilmi\u015f CUDA \u00c7ekirdekleri:<\/strong> vLLM, LLM operasyonlar\u0131 i\u00e7in \u00f6zel olarak optimize edilmi\u015f CUDA \u00e7ekirdekleri kullan\u0131r. Bu \u00e7ekirdekler, modelin hesaplama yo\u011fun k\u0131s\u0131mlar\u0131n\u0131 daha h\u0131zl\u0131 \u00e7al\u0131\u015ft\u0131rarak genel performans\u0131 art\u0131r\u0131r.<\/li>\n<li><strong>\u00c7oklu GPU Deste\u011fi:<\/strong> B\u00fcy\u00fck modellerin birden fazla GPU&#8217;ya yay\u0131lmas\u0131na olanak tan\u0131r, bu da daha b\u00fcy\u00fck modellerin veya daha y\u00fcksek y\u00fcklerin i\u015flenmesini sa\u011flar.<\/li>\n<\/ul>\n<p>Bu optimizasyonlar sayesinde vLLM, Hugging Face Transformers k\u00fct\u00fcphanesinin standart \u00e7\u0131kar\u0131m boru hatlar\u0131na g\u00f6re 24 kata kadar daha y\u00fcksek verim sa\u011flayabilir. Bu, DeepScaleR-1.5B gibi modelleri DigitalOcean&#8217;da maliyet etkin bir \u015fekilde bar\u0131nd\u0131rmak i\u00e7in vLLM&#8217;i ideal bir se\u00e7im haline getirir.<\/p>\n<h3>DigitalOcean: Bulut Ortam\u0131 Se\u00e7imi<\/h3>\n<p>DigitalOcean, geli\u015ftiricilere ve k\u00fc\u00e7\u00fck\/orta \u00f6l\u00e7ekli i\u015fletmelere y\u00f6nelik basit, uygun maliyetli ve \u00f6l\u00e7eklenebilir bulut hizmetleri sunan pop\u00fcler bir platformdur. LLM da\u011f\u0131t\u0131mlar\u0131 i\u00e7in DigitalOcean&#8217;\u0131 tercih etmemizin ba\u015fl\u0131ca nedenleri \u015funlard\u0131r:<\/p>\n<ul>\n<li><strong>Kullan\u0131m Kolayl\u0131\u011f\u0131:<\/strong> DigitalOcean&#8217;\u0131n kullan\u0131c\u0131 dostu aray\u00fcz\u00fc ve basit API&#8217;leri, altyap\u0131 kurulumunu ve y\u00f6netimini kolayla\u015ft\u0131r\u0131r.<\/li>\n<li><strong>Maliyet Etkinli\u011fi:<\/strong> Di\u011fer b\u00fcy\u00fck bulut sa\u011flay\u0131c\u0131lar\u0131na k\u0131yasla rekabet\u00e7i fiyatland\u0131rma sunar, \u00f6zellikle GPU kaynaklar\u0131 i\u00e7in daha ula\u015f\u0131labilir se\u00e7enekler sunabilir.<\/li>\n<li><strong>GPU Droplet&#8217;ler:<\/strong> LLM&#8217;lerin \u00e7al\u0131\u015fmas\u0131 i\u00e7in vazge\u00e7ilmez olan NVIDIA GPU&#8217;lara sahip &#8220;Droplet&#8221; (sanal makine) se\u00e7enekleri mevcuttur. Bu droplet&#8217;ler genellikle NVIDIA A100 veya A6000 gibi y\u00fcksek performansl\u0131 GPU&#8217;lar\u0131 bar\u0131nd\u0131r\u0131r, bu da DeepScaleR-1.5B gibi modeller i\u00e7in yeterli hesaplama g\u00fcc\u00fc sa\u011flar.<\/li>\n<li><strong>\u00d6l\u00e7eklenebilirlik:<\/strong> \u0130htiya\u00e7 halinde daha fazla kaynak eklemek veya farkl\u0131 b\u00f6lgelerde da\u011f\u0131t\u0131m yapmak m\u00fcmk\u00fcnd\u00fcr.<\/li>\n<li><strong>Geli\u015ftirici Dostu:<\/strong> Sa\u011flad\u0131\u011f\u0131 ara\u00e7lar ve dok\u00fcmantasyon, geli\u015ftiricilerin h\u0131zl\u0131 bir \u015fekilde ba\u015flamas\u0131na olanak tan\u0131r.<\/li>\n<\/ul>\n<p>Bu makalede, DeepScaleR-1.5B modelini vLLM ile DigitalOcean&#8217;\u0131n GPU destekli bir Droplet&#8217;i \u00fczerinde \u00e7al\u0131\u015ft\u0131raca\u011f\u0131z.<\/p>\n<h3>Kurulum \u00d6ncesi Haz\u0131rl\u0131klar<\/h3>\n<p>Da\u011f\u0131t\u0131m s\u00fcrecine ba\u015flamadan \u00f6nce baz\u0131 temel haz\u0131rl\u0131klar\u0131 yapmam\u0131z gerekmektedir.<\/p>\n<h4>DigitalOcean Hesab\u0131 ve GPU Droplet Olu\u015fturma<\/h4>\n<ol>\n<li><strong>DigitalOcean Hesab\u0131:<\/strong> E\u011fer bir hesab\u0131n\u0131z yoksa, DigitalOcean web sitesi \u00fczerinden bir hesap olu\u015fturun.<\/li>\n<li><strong>GPU Droplet Se\u00e7imi:<\/strong> DigitalOcean kontrol panelinden &#8220;Create Droplet&#8221; se\u00e7ene\u011fini t\u0131klay\u0131n.\n<ul>\n<li><strong>B\u00f6lge (Region):<\/strong> GPU Droplet&#8217;lerinin bulundu\u011fu bir b\u00f6lge se\u00e7in. Genellikle ABD (New York, San Francisco) veya Avrupa (Frankfurt) b\u00f6lgelerinde GPU se\u00e7enekleri daha yayg\u0131nd\u0131r.<\/li>\n<li><strong>\u0130\u015fletim Sistemi (OS):<\/strong> Ubuntu Server (tercihen en son LTS s\u00fcr\u00fcm\u00fc, \u00f6rn. 22.04 LTS) se\u00e7in.<\/li>\n<li><strong>Droplet Tipi:<\/strong> &#8220;Premium CPU-Optimized&#8221; veya &#8220;General Purpose&#8221; alt\u0131nda GPU se\u00e7eneklerini aray\u0131n. NVIDIA A100 veya A6000 gibi GPU&#8217;lara sahip bir Droplet se\u00e7in. Modelin boyutu ve beklenen y\u00fck g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, en az 1x A100 (80GB VRAM) veya 1x A6000 (48GB VRAM) gibi bir GPU yeterli olacakt\u0131r. Daha d\u00fc\u015f\u00fck kapasiteli GPU&#8217;lar da DeepScaleR-1.5B i\u00e7in uygun olabilir, ancak performans beklentilerinizi buna g\u00f6re ayarlamal\u0131s\u0131n\u0131z.<\/li>\n<li><strong>CPU ve RAM:<\/strong> GPU&#8217;lu Droplet&#8217;ler genellikle yeterli CPU ve RAM ile gelir. Ancak, en az 8 vCPU ve 32 GB RAM \u00f6nerilir.<\/li>\n<li><strong>Depolama (Storage):<\/strong> Model a\u011f\u0131rl\u0131klar\u0131n\u0131n boyutu ve loglar i\u00e7in yeterli disk alan\u0131 se\u00e7in. En az 100-200 GB SSD \u00f6nerilir.<\/li>\n<li><strong>SSH Anahtarlar\u0131:<\/strong> SSH ile g\u00fcvenli ba\u011flant\u0131 i\u00e7in bir SSH anahtar\u0131 ekleyin veya olu\u015fturun. Bu, Droplet&#8217;e \u015fifre yerine anahtar ile eri\u015fmenizi sa\u011flar.<\/li>\n<li><strong>G\u00fcvenlik Duvar\u0131 (Firewall):<\/strong> Bir g\u00fcvenlik duvar\u0131 olu\u015fturarak SSH (port 22) ve vLLM API&#8217;sinin \u00e7al\u0131\u015faca\u011f\u0131 portu (varsay\u0131lan 8000) a\u00e7\u0131n.<\/li>\n<\/ul>\n<\/li>\n<li>Droplet&#8217;i olu\u015fturduktan sonra, genel IP adresini not al\u0131n.<\/li>\n<\/ol>\n<h4>SSH ile Ba\u011flant\u0131 ve Temel Sistem G\u00fcncellemeleri<\/h4>\n<p>Droplet&#8217;in genel IP adresi ve SSH anahtar\u0131n\u0131zla sunucuya ba\u011flan\u0131n:<\/p>\n<pre><code class=\"language-bash\">ssh -i \/yol\/to\/anahtar.pem root@<Droplet_IP_Adresi><\/code><\/pre>\n<p>Ba\u011fland\u0131ktan sonra sistemi g\u00fcncelleyin ve temel paketleri kurun:<\/p>\n<pre><code class=\"language-bash\">sudo apt update && sudo apt upgrade -y\nsudo apt install -y build-essential git python3-pip python3-venv<\/code><\/pre>\n<h4>NVIDIA S\u00fcr\u00fcc\u00fcleri ve CUDA Toolkit Kurulumu<\/h4>\n<p>Bu ad\u0131m, LLM&#8217;lerin GPU \u00fczerinde \u00e7al\u0131\u015fmas\u0131 i\u00e7in kritik \u00f6neme sahiptir.<\/p>\n<ol>\n<li><strong>NVIDIA S\u00fcr\u00fcc\u00fcleri:<\/strong> DigitalOcean&#8217;\u0131n GPU Droplet&#8217;leri genellikle \u00f6nceden y\u00fcklenmi\u015f NVIDIA s\u00fcr\u00fcc\u00fcleriyle gelir. Ancak, en g\u00fcncel ve uyumlu s\u00fcr\u00fcc\u00fcleri kurmak her zaman iyi bir uygulamad\u0131r. NVIDIA&#8217;n\u0131n resmi web sitesinden veya Ubuntu&#8217;nun <code>apt<\/code> deposundan s\u00fcr\u00fcc\u00fcleri kurabilirsiniz.\n<pre><code class=\"language-bash\"># Mevcut s\u00fcr\u00fcc\u00fcleri kontrol et\n        nvidia-smi\n\n        # E\u011fer s\u00fcr\u00fcc\u00fcler kurulu de\u011filse veya g\u00fcncel de\u011filse\n        sudo apt install -y nvidia-driver-535 # Veya en g\u00fcncel kararl\u0131 s\u00fcr\u00fcm\n        sudo reboot # S\u00fcr\u00fcc\u00fc kurulumundan sonra yeniden ba\u015flatma gerekebilir<\/code><\/pre>\n<p>        Yeniden ba\u015flatt\u0131ktan sonra <code>nvidia-smi<\/code> komutunu tekrar \u00e7al\u0131\u015ft\u0131rarak s\u00fcr\u00fcc\u00fclerin do\u011fru bir \u015fekilde y\u00fcklendi\u011fini do\u011frulay\u0131n.<\/li>\n<li><strong>CUDA Toolkit:<\/strong> vLLM ve PyTorch gibi k\u00fct\u00fcphaneler i\u00e7in CUDA Toolkit gereklidir. NVIDIA&#8217;n\u0131n geli\u015ftirici web sitesinden uygun s\u00fcr\u00fcm\u00fc indirin ve kurun. Genellikle PyTorch&#8217;un destekledi\u011fi en son kararl\u0131 CUDA s\u00fcr\u00fcm\u00fcn\u00fc (\u00f6rn. CUDA 11.8 veya 12.1) tercih edin.\n<pre><code class=\"language-bash\"># \u00d6rnek olarak CUDA 11.8 kurulumu (resmi NVIDIA talimatlar\u0131na g\u00f6re de\u011fi\u015febilir)\n        wget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu2204\/x86_64\/cuda-ubuntu2204.pin\n        sudo mv cuda-ubuntu2204.pin \/etc\/apt\/preferences.d\/cuda-repository-pin-600\n        wget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu2204\/x86_64\/cuda-repo-ubuntu2204-11-8-local_11.8.0-1_amd64.deb\n        sudo dpkg -i cuda-repo-ubuntu2204-11-8-local_11.8.0-1_amd64.deb\n        sudo cp \/var\/cuda-repo-ubuntu2204-11-8-local\/cuda-*-keyring.gpg \/usr\/share\/keyrings\/\n        sudo apt update\n        sudo apt install -y cuda-toolkit-11-8\n\n        # Ortam de\u011fi\u015fkenlerini ayarlay\u0131n\n        echo 'export PATH=\/usr\/local\/cuda-11.8\/bin:$PATH' >> ~\/.bashrc\n        echo 'export LD_LIBRARY_PATH=\/usr\/local\/cuda-11.8\/lib64:$LD_LIBRARY_PATH' >> ~\/.bashrc\n        source ~\/.bashrc<\/code><\/pre>\n<p>        <code>nvcc --version<\/code> komutu ile CUDA kurulumunu do\u011frulay\u0131n.<\/li>\n<\/ol>\n<h3>DeepScaleR-1.5B Modelini Haz\u0131rlama<\/h3>\n<p>DeepScaleR-1.5B model a\u011f\u0131rl\u0131klar\u0131n\u0131 sunucuya indirmemiz gerekmektedir. Bu modelin Hugging Face Hub \u00fczerinde <code>DeepScaleAI\/DeepScaleR-1.5B<\/code> ad\u0131yla bulundu\u011funu varsayal\u0131m.<\/p>\n<h4>Python Sanal Ortam\u0131 Olu\u015fturma ve K\u00fct\u00fcphaneleri Kurma<\/h4>\n<p>Model ve vLLM i\u00e7in izole bir Python ortam\u0131 olu\u015ftural\u0131m:<\/p>\n<pre><code class=\"language-bash\">python3 -m venv vllm_env\nsource vllm_env\/bin\/activate<\/code><\/pre>\n<p>\u015eimdi gerekli Python k\u00fct\u00fcphanelerini kural\u0131m. \u00d6zellikle PyTorch&#8217;u CUDA deste\u011fiyle kurmak \u00f6nemlidir. CUDA s\u00fcr\u00fcm\u00fcn\u00fczle uyumlu PyTorch s\u00fcr\u00fcm\u00fcn\u00fc se\u00e7in (\u00f6rne\u011fin, CUDA 11.8 i\u00e7in <code>cu118<\/code>).<\/p>\n<pre><code class=\"language-bash\">pip install --upgrade pip\npip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https:\/\/download.pytorch.org\/whl\/cu118\npip install transformers==4.35.2 accelerate==0.24.1<\/code><\/pre>\n<p>Bu komutlar, PyTorch&#8217;un CUDA 11.8 ile uyumlu s\u00fcr\u00fcm\u00fcn\u00fc ve Transformers k\u00fct\u00fcphanesinin belirtilen s\u00fcr\u00fcmlerini kurar.<\/p>\n<h4>Model A\u011f\u0131rl\u0131klar\u0131n\u0131 \u0130ndirme<\/h4>\n<p>Hugging Face Hub&#8217;dan DeepScaleR-1.5B model a\u011f\u0131rl\u0131klar\u0131n\u0131 indirmek i\u00e7in <code>transformers<\/code> k\u00fct\u00fcphanesini kullanabiliriz veya direkt <code>huggingface-cli<\/code> arac\u0131n\u0131 kullanabiliriz.<\/p>\n<pre><code class=\"language-bash\"># Model a\u011f\u0131rl\u0131klar\u0131n\u0131 depolamak i\u00e7in bir dizin olu\u015fturun\nmkdir -p models\/DeepScaleR-1.5B\ncd models\/DeepScaleR-1.5B\n\n<h2>Model a\u011f\u0131rl\u0131klar\u0131n\u0131 indirin (bu biraz zaman alabilir)<\/h2>\n<h2>transformers k\u00fct\u00fcphanesi ile:<\/h2>\npython -c \"from transformers import AutoTokenizer, AutoModelForCausalLM; AutoTokenizer.from_pretrained('DeepScaleAI\/DeepScaleR-1.5B'); AutoModelForCausalLM.from_pretrained('DeepScaleAI\/DeepScaleR-1.5B')\"\n\n<h2>Alternatif olarak, huggingface-cli ile:<\/h2>\n<h2>pip install huggingface_hub<\/h2>\n<h2>huggingface-cli download DeepScaleAI\/DeepScaleR-1.5B --local-dir . --local-dir-use-symlinks False<\/code><\/pre>\n<\/h2>\n<p>Bu i\u015flem, modelin t\u00fcm a\u011f\u0131rl\u0131klar\u0131n\u0131 ve konfig\u00fcrasyon dosyalar\u0131n\u0131 <code>models\/DeepScaleR-1.5B<\/code> dizinine indirecektir.<\/p>\n<h3>vLLM Kurulumu ve Yap\u0131land\u0131rmas\u0131<\/h3>\n<p>Model a\u011f\u0131rl\u0131klar\u0131 haz\u0131r oldu\u011funa g\u00f6re, vLLM&#8217;i kurabilir ve DeepScaleR-1.5B ile bir \u00e7\u0131kar\u0131m sunucusu ba\u015flatabiliriz.<\/p>\n<h4>vLLM Kurulumu<\/h4>\n<p>Aktif Python sanal ortam\u0131n\u0131zda vLLM&#8217;i kurun:<\/p>\n<pre><code class=\"language-bash\">pip install vllm==0.2.7 # En son kararl\u0131 s\u00fcr\u00fcm\u00fc kontrol edin ve kurun<\/code><\/pre>\n<p>vLLM, gerekli CUDA uzant\u0131lar\u0131n\u0131 derleyece\u011fi i\u00e7in bu i\u015flem biraz zaman alabilir.<\/p>\n<h4>vLLM Sunucusunu Ba\u015flatma<\/h4>\n<p>\u015eimdi DeepScaleR-1.5B modelini vLLM ile \u00e7al\u0131\u015ft\u0131rmak i\u00e7in bir sunucu ba\u015flatabiliriz. Model dizininin do\u011fru oldu\u011fundan emin olun.<\/p>\n<pre><code class=\"language-bash\">python -m vllm.entrypoints.api_server \\\n    --model \/root\/models\/DeepScaleR-1.5B \\\n    --host 0.0.0.0 \\\n    --port 8000 \\\n    --tensor-parallel-size 1 \\\n    --gpu-memory-utilization 0.9 \\\n    --dtype bfloat16 \\\n    --max-model-len 2048 \\\n    --enforce-eager<\/code><\/pre>\n<p>Bu komutun parametreleri \u015funlard\u0131r:<\/p>\n<ul>\n<li><code>--model \/root\/models\/DeepScaleR-1.5B<\/code>: DeepScaleR-1.5B modelinin yerel yolu.<\/li>\n<li><code>--host 0.0.0.0<\/code>: API&#8217;nin t\u00fcm a\u011f aray\u00fczlerinden eri\u015filebilir olmas\u0131n\u0131 sa\u011flar.<\/li>\n<li><code>--port 8000<\/code>: API&#8217;nin dinleyece\u011fi port numaras\u0131. DigitalOcean g\u00fcvenlik duvar\u0131n\u0131zda bu portu a\u00e7t\u0131\u011f\u0131n\u0131zdan emin olun.<\/li>\n<li><code>--tensor-parallel-size 1<\/code>: Modelin ka\u00e7 GPU&#8217;ya yay\u0131laca\u011f\u0131n\u0131 belirtir. Tek GPU&#8217;lu bir Droplet kulland\u0131\u011f\u0131m\u0131z i\u00e7in 1 olarak ayarlanm\u0131\u015ft\u0131r. Birden fazla GPU&#8217;nuz varsa ve model \u00e7ok b\u00fcy\u00fckse bu de\u011feri art\u0131rabilirsiniz.<\/li>\n<li><code>--gpu-memory-utilization 0.9<\/code>: GPU belle\u011finin y\u00fczde ka\u00e7\u0131n\u0131n KV \u00f6nbelle\u011fi i\u00e7in kullan\u0131laca\u011f\u0131n\u0131 belirler. Kalan k\u0131s\u0131m model a\u011f\u0131rl\u0131klar\u0131 ve di\u011fer operasyonlar i\u00e7in ayr\u0131l\u0131r. %90, genellikle iyi bir ba\u015flang\u0131\u00e7 noktas\u0131d\u0131r.<\/li>\n<li><code>--dtype bfloat16<\/code>: Modelin \u00e7\u0131kar\u0131m i\u00e7in kullanaca\u011f\u0131 veri tipini belirtir. <code>bfloat16<\/code> veya <code>float16<\/code> (FP16) genellikle daha h\u0131zl\u0131 ve daha az bellek t\u00fcketirken, <code>float32<\/code> (FP32) daha y\u00fcksek do\u011fruluk sunar ancak daha yava\u015ft\u0131r ve daha fazla bellek kullan\u0131r. \u00c7o\u011fu modern GPU <code>bfloat16<\/code>&#8216;y\u0131 destekler ve LLM&#8217;ler i\u00e7in iyi bir denge sunar.<\/li>\n<li><code>--max-model-len 2048<\/code>: Modelin i\u015fleyebilece\u011fi maksimum ba\u011flam uzunlu\u011fu (input token + output token). Modelin orijinal ba\u011flam uzunlu\u011funa g\u00f6re ayarlay\u0131n.<\/li>\n<li><code>--enforce-eager<\/code>: PyTorch&#8217;un eager modunu kullanmaya zorlar. Baz\u0131 durumlarda hata ay\u0131klamay\u0131 kolayla\u015ft\u0131rabilir. Performans i\u00e7in kald\u0131r\u0131labilir.<\/li>\n<\/ul>\n<p>Komutu \u00e7al\u0131\u015ft\u0131rd\u0131\u011f\u0131n\u0131zda, vLLM model a\u011f\u0131rl\u0131klar\u0131n\u0131 GPU belle\u011fine y\u00fckleyecek ve API sunucusunu ba\u015flatacakt\u0131r. Konsolda &#8220;Uvicorn running on http:\/\/0.0.0.0:8000&#8221; gibi bir mesaj g\u00f6rmelisiniz.<\/p>\n<h4>Arka Planda S\u00fcrekli \u00c7al\u0131\u015ft\u0131rma (Opsiyonel)<\/h4>\n<p>SSH oturumunuzu kapatt\u0131\u011f\u0131n\u0131zda sunucunun durmas\u0131n\u0131 engellemek i\u00e7in <code>nohup<\/code> kullanarak veya bir <code>systemd<\/code> servisi olarak \u00e7al\u0131\u015ft\u0131rabilirsiniz.<\/p>\n<p><strong><code>nohup<\/code> ile:<\/strong><\/p>\n<pre><code class=\"language-bash\">nohup python -m vllm.entrypoints.api_server \\\n    --model \/root\/models\/DeepScaleR-1.5B \\\n    --host 0.0.0.0 \\\n    --port 8000 \\\n    --tensor-parallel-size 1 \\\n    --gpu-memory-utilization 0.9 \\\n    --dtype bfloat16 \\\n    --max-model-len 2048 \\\n    --enforce-eager > vllm_server.log 2>&1 &<\/code><\/pre>\n<p>Bu komut, sunucuyu arka planda \u00e7al\u0131\u015ft\u0131r\u0131r ve t\u00fcm \u00e7\u0131kt\u0131lar\u0131 <code>vllm_server.log<\/code> dosyas\u0131na y\u00f6nlendirir.<\/p>\n<p><strong><code>systemd<\/code> Servisi Olarak:<\/strong><br \/>\nDaha sa\u011flam bir \u00e7\u00f6z\u00fcm i\u00e7in bir <code>systemd<\/code> servisi olu\u015fturabilirsiniz.<\/p>\n<pre><code class=\"language-bash\">sudo nano \/etc\/systemd\/system\/vllm.service<\/code><\/pre>\n<p>Dosyan\u0131n i\u00e7ine a\u015fa\u011f\u0131daki i\u00e7eri\u011fi yap\u0131\u015ft\u0131r\u0131n (yollar\u0131 ve parametreleri kendi kurulumunuza g\u00f6re ayarlay\u0131n):<\/p>\n<pre><code class=\"language-ini\">[Unit]\nDescription=vLLM API Server for DeepScaleR-1.5B\nAfter=network.target\n\n[Service]\nUser=root # veya kendi kullan\u0131c\u0131 ad\u0131n\u0131z\nWorkingDirectory=\/root\nExecStart=\/root\/vllm_env\/bin\/python -m vllm.entrypoints.api_server \\\n    --model \/root\/models\/DeepScaleR-1.5B \\\n    --host 0.0.0.0 \\\n    --port 8000 \\\n    --tensor-parallel-size 1 \\\n    --gpu-memory-utilization 0.9 \\\n    --dtype bfloat16 \\\n    --max-model-len 2048 \\\n    --enforce-eager\nRestart=always\nStandardOutput=journal\nStandardError=journal\n\n[Install]\nWantedBy=multi-user.target<\/code><\/pre>\n<p>Kaydedip kapat\u0131n, ard\u0131ndan servisi etkinle\u015ftirin ve ba\u015flat\u0131n:<\/p>\n<pre><code class=\"language-bash\">sudo systemctl daemon-reload\nsudo systemctl enable vllm.service\nsudo systemctl start vllm.service\nsudo systemctl status vllm.service<\/code><\/pre>\n<p>Servisin durumunu kontrol etmek i\u00e7in <code>sudo systemctl status vllm.service<\/code> komutunu kullanabilirsiniz. Loglar\u0131 g\u00f6rmek i\u00e7in <code>journalctl -u vllm.service<\/code> komutu i\u015fe yarar.<\/p>\n<h3>API Testi ve Kullan\u0131m<\/h3>\n<p>vLLM sunucusu \u00e7al\u0131\u015f\u0131r duruma geldi\u011finde, Droplet&#8217;inizin genel IP adresi ve 8000 numaral\u0131 port \u00fczerinden API&#8217;ye eri\u015febilirsiniz.<\/p>\n<h4><code>curl<\/code> ile API Testi<\/h4>\n<p>Yerel makinenizden veya ba\u015fka bir sunucudan a\u015fa\u011f\u0131daki <code>curl<\/code> komutunu kullanarak bir istek g\u00f6nderebilirsiniz:<\/p>\n<pre><code class=\"language-bash\">curl http:\/\/<Droplet_IP_Adresi>:8000\/generate \\\n    -H \"Content-Type: application\/json\" \\\n    -d '{\n        \"prompt\": \"B\u00fcy\u00fck dil modellerinin gelece\u011fi nedir?\",\n        \"max_tokens\": 100,\n        \"temperature\": 0.7,\n        \"top_p\": 0.9,\n        \"n\": 1,\n        \"stream\": false\n    }'<\/code><\/pre>\n<p>Bu komut, DeepScaleR-1.5B modelinden &#8220;B\u00fcy\u00fck dil modellerinin gelece\u011fi nedir?&#8221; sorusuna 100 token&#8217;a kadar bir yan\u0131t \u00fcretmesini ister. Yan\u0131t JSON format\u0131nda gelecektir.<\/p>\n<h4>Python ile \u0130stemci Olu\u015fturma<\/h4>\n<p>Bir Python istemcisi kullanarak vLLM API&#8217;sine programatik olarak eri\u015fmek daha yayg\u0131n bir yakla\u015f\u0131md\u0131r:<\/p>\n<pre><code class=\"language-python\">import requests\nimport json\n\napi_url = \"http:\/\/<Droplet_IP_Adresi>:8000\/generate\"\nheaders = {\"Content-Type\": \"application\/json\"}\n\ndata = {\n    \"prompt\": \"DeepScaleR-1.5B modelinin DigitalOcean \u00fczerindeki avantajlar\u0131n\u0131 a\u00e7\u0131kla.\",\n    \"max_tokens\": 150,\n    \"temperature\": 0.7,\n    \"top_p\": 0.9,\n    \"n\": 1,\n    \"stream\": False # True yaparsan\u0131z ak\u0131\u015fl\u0131 yan\u0131t al\u0131rs\u0131n\u0131z\n}\n\ntry:\n    response = requests.post(api_url, headers=headers, data=json.dumps(data))\n    response.raise_for_status() # HTTP hatalar\u0131n\u0131 kontrol et\n    result = response.json()\n    \n    if \"text\" in result[\"choices\"][0]:\n        print(\"Model Yan\u0131t\u0131:\", result[\"choices\"][0][\"text\"])\n    else:\n        print(\"Yan\u0131t format\u0131 beklenenden farkl\u0131:\", result)\n\nexcept requests.exceptions.RequestException as e:\n    print(f\"API iste\u011fi s\u0131ras\u0131nda hata olu\u015ftu: {e}\")\nexcept json.JSONDecodeError:\n    print(\"API'den ge\u00e7ersiz JSON yan\u0131t\u0131 al\u0131nd\u0131.\")<\/code><\/pre>\n<p>Bu \u00f6rnek kod, Python <code>requests<\/code> k\u00fct\u00fcphanesini kullanarak vLLM sunucusuna bir POST iste\u011fi g\u00f6nderir ve modelin yan\u0131t\u0131n\u0131 yazd\u0131r\u0131r. <code>stream=True<\/code> olarak ayarlarsan\u0131z, yan\u0131tlar\u0131 par\u00e7a par\u00e7a i\u015flemek i\u00e7in ak\u0131\u015fl\u0131 okuma yapman\u0131z gerekir.<\/p>\n<h3>Performans \u0130zleme ve Optimizasyon<\/h3>\n<p>LLM hizmetinizin performans\u0131 ve maliyet etkinli\u011fi i\u00e7in izleme ve optimizasyon \u00f6nemlidir.<\/p>\n<h4>Performans \u0130zleme Ara\u00e7lar\u0131<\/h4>\n<ul>\n<li><strong><code>nvidia-smi<\/code>:<\/strong> Droplet \u00fczerinde <code>nvidia-smi<\/code> komutunu \u00e7al\u0131\u015ft\u0131rarak GPU kullan\u0131m\u0131n\u0131 (bellek, GPU \u00e7ekirdek kullan\u0131m\u0131, s\u0131cakl\u0131k) anl\u0131k olarak izleyebilirsiniz.<\/li>\n<li><strong>vLLM Loglar\u0131:<\/strong> vLLM sunucusunun \u00e7\u0131kt\u0131s\u0131n\u0131 y\u00f6nlendirdi\u011finiz <code>vllm_server.log<\/code> dosyas\u0131n\u0131 (<code>tail -f vllm_server.log<\/code>) veya <code>journalctl -u vllm.service<\/code> komutunu izleyerek gelen istekleri, \u00e7\u0131kar\u0131m s\u00fcrelerini ve olas\u0131 hatalar\u0131 g\u00f6rebilirsiniz.<\/li>\n<li><strong>DigitalOcean Metrikleri:<\/strong> DigitalOcean kontrol panelindeki Droplet metrikleri (CPU kullan\u0131m\u0131, RAM kullan\u0131m\u0131, a\u011f trafi\u011fi) sunucunuzun genel durumunu anlaman\u0131za yard\u0131mc\u0131 olur.<\/li>\n<\/ul>\n<h4>Optimizasyon \u0130pu\u00e7lar\u0131<\/h4>\n<ul>\n<li><strong><code>--gpu-memory-utilization<\/code>:<\/strong> Bu parametreyi dikkatli ayarlay\u0131n. \u00c7ok y\u00fcksek bir de\u011fer, model a\u011f\u0131rl\u0131klar\u0131 i\u00e7in yeterli bellek b\u0131rakmayabilir ve &#8220;out of memory&#8221; hatalar\u0131na yol a\u00e7abilir. \u00c7ok d\u00fc\u015f\u00fck bir de\u011fer ise KV \u00f6nbellek alan\u0131n\u0131 k\u0131s\u0131tlayarak verimi d\u00fc\u015f\u00fcrebilir. Deneyerek en uygun de\u011feri bulun.<\/li>\n<li><strong><code>--dtype<\/code>:<\/strong> <code>bfloat16<\/code> veya <code>float16<\/code> (FP16) kullanmak, bellek t\u00fcketimini ve hesaplama s\u00fcresini \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r. \u00c7o\u011fu LLM i\u00e7in <code>bfloat16<\/code> yeterli do\u011fruluk sa\u011flar.<\/li>\n<li><strong><code>--max-model-len<\/code>:<\/strong> Modelin destekledi\u011fi maksimum ba\u011flam uzunlu\u011funu a\u015fmay\u0131n. Bu de\u011feri ihtiyac\u0131n\u0131za g\u00f6re ayarlay\u0131n. \u00c7ok y\u00fcksek bir de\u011fer, daha fazla KV \u00f6nbellek alan\u0131 ay\u0131rarak di\u011fer istekler i\u00e7in alan\u0131 k\u0131s\u0131tlar.<\/li>\n<li><strong>Gruplama (Batching):<\/strong> vLLM&#8217;in s\u00fcrekli gruplama \u00f6zelli\u011fi sayesinde, istemci taraf\u0131nda \u00f6zel bir gruplama yapman\u0131za gerek kalmaz. Ancak, e\u015fzamanl\u0131 istek say\u0131s\u0131n\u0131 art\u0131rarak vLLM&#8217;in verimini art\u0131rabilirsiniz.<\/li>\n<li><strong>Model Kuantizasyonu:<\/strong> E\u011fer DeepScaleR-1.5B&#8217;nin kuantize edilmi\u015f (\u00f6rn. 8-bit, 4-bit) versiyonlar\u0131 mevcutsa, bunlar\u0131 kullanarak bellek t\u00fcketimini daha da d\u00fc\u015f\u00fcrebilir ve performans\u0131 art\u0131rabilirsiniz. vLLM, bu t\u00fcr kuantize edilmi\u015f modelleri de destekler.<\/li>\n<li><strong>Maliyet Y\u00f6netimi:<\/strong> Kullanmad\u0131\u011f\u0131n\u0131z zamanlarda Droplet&#8217;i kapatarak (power off) maliyetlerden tasarruf edebilirsiniz. Ancak, model a\u011f\u0131rl\u0131klar\u0131 bellekte kalmayaca\u011f\u0131ndan tekrar ba\u015flatma s\u00fcresi uzayacakt\u0131r. Otomatik \u00f6l\u00e7eklendirme veya DigitalOcean App Platform gibi hizmetleri kullanarak daha dinamik \u00e7\u00f6z\u00fcmler d\u00fc\u015f\u00fcnebilirsiniz.<\/li>\n<\/ul>\n<h3>G\u00fcvenlik ve Bak\u0131m<\/h3>\n<p>\u00dcretim ortam\u0131ndaki bir LLM hizmetinin g\u00fcvenli\u011fi ve d\u00fczenli bak\u0131m\u0131 hayati \u00f6nem ta\u015f\u0131r.<\/p>\n<ul>\n<li><strong>SSH Anahtar\u0131 G\u00fcvenli\u011fi:<\/strong> SSH anahtar\u0131n\u0131z\u0131 asla payla\u015fmay\u0131n ve g\u00fcvenli bir yerde saklay\u0131n.<\/li>\n<li><strong>G\u00fcvenlik Duvar\u0131 Kurallar\u0131:<\/strong> Sadece gerekli portlar\u0131 (SSH i\u00e7in 22, vLLM API i\u00e7in 8000) a\u00e7\u0131n. Di\u011fer t\u00fcm portlar\u0131 kapat\u0131n. \u0130stemci IP&#8217;lerini k\u0131s\u0131tlayarak eri\u015fimi daha da daraltabilirsiniz.<\/li>\n<li><strong>Sistem ve Yaz\u0131l\u0131m G\u00fcncellemeleri:<\/strong> \u0130\u015fletim sisteminizi ve t\u00fcm y\u00fckl\u00fc k\u00fct\u00fcphaneleri (PyTorch, Transformers, vLLM) d\u00fczenli olarak g\u00fcncelleyin. Bu, g\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131 kapat\u0131r ve performans iyile\u015ftirmelerinden faydalanman\u0131z\u0131 sa\u011flar.<\/li>\n<li><strong>Yedekleme:<\/strong> Model a\u011f\u0131rl\u0131klar\u0131n\u0131z\u0131n ve \u00f6nemli konfig\u00fcrasyon dosyalar\u0131n\u0131z\u0131n yedeklerini al\u0131n. DigitalOcean&#8217;\u0131n Snapshot veya Volume hizmetlerini kullanarak Droplet&#8217;inizin veya verilerinizin yedeklerini d\u00fczenli olarak olu\u015fturabilirsiniz.<\/li>\n<li><strong>Eri\u015fim Kontrol\u00fc:<\/strong> API&#8217;nizi internete a\u00e7\u0131yorsan\u0131z, bir API Gateway veya ters proxy (\u00f6rn. Nginx) kullanarak kimlik do\u011frulama, h\u0131z s\u0131n\u0131rlama ve SSL\/TLS \u015fifrelemesi eklemeyi d\u00fc\u015f\u00fcnebilirsiniz. Bu, do\u011frudan vLLM sunucusuna eri\u015fimi engelleyerek ek bir g\u00fcvenlik katman\u0131 sa\u011flar.<\/li>\n<\/ul>\n<h3>Sonu\u00e7<\/h3>\n<p>Bu makalede, DeepScaleR-1.5B gibi optimize edilmi\u015f bir B\u00fcy\u00fck Dil Modelini, vLLM \u00e7\u0131kar\u0131m \u00e7er\u00e7evesinin sundu\u011fu \u00fcst\u00fcn performans optimizasyonlar\u0131yla DigitalOcean bulut platformunda nas\u0131l da\u011f\u0131taca\u011f\u0131m\u0131z\u0131 ve \u00e7al\u0131\u015ft\u0131raca\u011f\u0131m\u0131z\u0131 ad\u0131m ad\u0131m inceledik. DigitalOcean&#8217;\u0131n eri\u015filebilir GPU Droplet&#8217;leri ve vLLM&#8217;in PagedAttention gibi yenilik\u00e7i teknikleri sayesinde, maliyet etkin ve y\u00fcksek verimli bir LLM \u00e7\u0131kar\u0131m hizmeti kurmak m\u00fcmk\u00fcnd\u00fcr.<\/p>\n<p>Bu kurulum, geli\u015ftiricilere ve i\u015fletmelere, LLM&#8217;lerin g\u00fcc\u00fcnden faydalan\u0131rken ayn\u0131 zamanda altyap\u0131 maliyetlerini ve operasyonel karma\u015f\u0131kl\u0131\u011f\u0131 y\u00f6netme esnekli\u011fi sunar. DeepScaleR-1.5B gibi modeller, belirli kullan\u0131m durumlar\u0131 i\u00e7in b\u00fcy\u00fck modellere k\u0131yasla daha pratik ve s\u00fcrd\u00fcr\u00fclebilir bir \u00e7\u00f6z\u00fcm sunar. Gelecekte, daha da optimize edilmi\u015f modeller, geli\u015fmi\u015f \u00e7\u0131kar\u0131m \u00e7er\u00e7eveleri ve bulut sa\u011flay\u0131c\u0131lar\u0131n\u0131n sundu\u011fu yeni GPU se\u00e7enekleri ile LLM da\u011f\u0131t\u0131mlar\u0131 daha da kolayla\u015facak ve daha geni\u015f kitlelere ula\u015facakt\u0131r. Bu rehber, kendi LLM projelerinizi bulutta hayata ge\u00e7irmek i\u00e7in sa\u011flam bir temel olu\u015fturmaktad\u0131r.<\/p>\n<p><\/body><\/p>\n","protected":false},"excerpt":{"rendered":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber\nGiri\u015f: B\u00fcy\u00fck Dil Modellerinin Da\u011f\u0131t\u0131m\u0131 v","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":[1401],"tags":[],"class_list":{"0":"post-31185","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-git","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v20.5 (Yoast SEO v25.3.1) - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber - Kodlar\u0131n Gizemli D\u00fcnyas\u0131<\/title>\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\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\" \/>\n<meta property=\"og:locale\" content=\"tr_TR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber\" \/>\n<meta property=\"og:description\" content=\"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber Giri\u015f: B\u00fcy\u00fck Dil Modellerinin Da\u011f\u0131t\u0131m\u0131 v\" \/>\n<meta property=\"og:url\" content=\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\" \/>\n<meta property=\"og:site_name\" content=\"Kodlar\u0131n Gizemli D\u00fcnyas\u0131\" \/>\n<meta property=\"article:published_time\" content=\"2025-10-06T21:41:10+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=\"18 dakika\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\"},\"author\":{\"name\":\"Fatih Soysal\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"headline\":\"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber\",\"datePublished\":\"2025-10-06T21:41:10+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\"},\"wordCount\":3041,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"articleSection\":[\"Git\"],\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#respond\"]}],\"copyrightYear\":\"2025\",\"copyrightHolder\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#organization\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\",\"name\":\"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber - Kodlar\u0131n Gizemli D\u00fcnyas\u0131\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\"},\"datePublished\":\"2025-10-06T21:41:10+00:00\",\"breadcrumb\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#breadcrumb\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Anasayfa\",\"item\":\"https:\/\/fatihsoysal.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber\"}]},{\"@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 SEO Premium plugin. -->","yoast_head_json":{"title":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","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\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/","og_locale":"tr_TR","og_type":"article","og_title":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber","og_description":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber Giri\u015f: B\u00fcy\u00fck Dil Modellerinin Da\u011f\u0131t\u0131m\u0131 v","og_url":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/","og_site_name":"Kodlar\u0131n Gizemli D\u00fcnyas\u0131","article_published_time":"2025-10-06T21:41:10+00:00","author":"Fatih Soysal","twitter_card":"summary_large_image","twitter_misc":{"Yazan:":"Fatih Soysal","Tahmini okuma s\u00fcresi":"18 dakika"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#article","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/"},"author":{"name":"Fatih Soysal","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"headline":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber","datePublished":"2025-10-06T21:41:10+00:00","mainEntityOfPage":{"@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/"},"wordCount":3041,"commentCount":0,"publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"articleSection":["Git"],"inLanguage":"tr","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#respond"]}],"copyrightYear":"2025","copyrightHolder":{"@id":"https:\/\/fatihsoysal.com\/blog\/#organization"}},{"@type":"WebPage","@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/","url":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/","name":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/#website"},"datePublished":"2025-10-06T21:41:10+00:00","breadcrumb":{"@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#breadcrumb"},"inLanguage":"tr","potentialAction":[{"@type":"ReadAction","target":["https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/fatihsoysal.com\/blog\/deepscaler-1-5b-modelini-vllm-ile-digitalocean-uzerinde-calistirma-kapsamli-bir-teknik-rehber\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Anasayfa","item":"https:\/\/fatihsoysal.com\/blog\/"},{"@type":"ListItem","position":2,"name":"DeepScaleR-1.5B Modelini vLLM ile DigitalOcean \u00dczerinde \u00c7al\u0131\u015ft\u0131rma: Kapsaml\u0131 Bir Teknik Rehber"}]},{"@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\/31185","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=31185"}],"version-history":[{"count":1,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/31185\/revisions"}],"predecessor-version":[{"id":31186,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/31185\/revisions\/31186"}],"wp:attachment":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/media?parent=31185"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/categories?post=31185"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/tags?post=31185"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}