ollama 설치

intel cpu 인데 내장 그래픽은 역시 버리는건가..(!)

 

$ curl -fsSL https://ollama.com/install.sh | sh >>> Installing ollama to /usr/local [sudo] minimonk 암호:  >>> Downloading ollama-linux-amd64.tar.zst ######################################################################## 100.0% >>> Creating ollama user... >>> Adding ollama user to render group... >>> Adding ollama user to video group... >>> Adding current user to ollama group... >>> Creating ollama systemd service... >>> Enabling and starting ollama service... Created symlink /etc/systemd/system/default.target.wants/ollama.service → /etc/systemd/system/ollama.service. >>> The Ollama API is now available at 127.0.0.1:11434. >>> Install complete. Run "ollama" from the command line. WARNING: No NVIDIA/AMD GPU detected. Ollama will run in CPU-only mode.

 

$ ollama  Ollama 0.20.3                                                                                               Chat with a model                                      Start an interactive chat with a model                                                                  Launch Claude Code (not installed)                     Anthropic's coding tool with subagents                                                                  Launch Codex (not installed)                           OpenAI's open-source coding agent                                                                       Launch OpenClaw (install)                              Personal AI with 100+ skills                                                                            Launch OpenCode (not installed)                        Anomaly's open-source coding agent                                                                      Launch Droid (not installed)                           Factory's coding agent across terminal and IDEs                                                       ▸ Launch Pi (install)                                    Press enter to install                                                                                                                                     ↑/↓ navigate • enter launch • → configure • esc quit

 

$ ollama  Select model to run: Type to filter...                                                                                                                              Recommended                                                                      ▸ kimi-k2.5:cloud                                                                    Multimodal reasoning with subagents                                            qwen3.5:cloud                                                                      Reasoning, coding, and agentic tool use with vision                            glm-5:cloud                                                                        Reasoning and code generation                                                  minimax-m2.7:cloud                                                                 Fast, efficient coding and real-world productivity                             gemma4                                                                             Reasoning and code generation locally, ~16GB, (not downloaded)                 qwen3.5                                                                            Reasoning, coding, and visual understanding locally, ~11GB, (not downloade                                                                                  ↑/↓ navigate • enter select • ← back

$ ollama  Download gemma4?                                                                       Yes    No                                                                        ←/→ navigate • enter confirm • esc cancel

 

메모리도 작고(16GB) cpu로 돌릴거라 취소!

pulling manifest  pulling 4c27e0f5b5ad:  40% ▕███████           ▏ 3.9 GB/9.6 GB   11 MB/s   8m11s

 

[링크 : https://goddaehee.tistory.com/]

[링크 : https://ollama.com/download]

[링크 : https://ollama.com/search]

 

+

일단 돌릴 테스트니까  1B던 3B던 상관없으니 시작

그런데 아까 위에 메뉴에서는 왜 gemma만 보이고 이제 llama는 안보이는걸까?

일단은.. updated 1 years ago 인거 봐서 최신으로만 보여주는걸려나?

$ ollama run llama3.2 pulling manifest  pulling dde5aa3fc5ff:  70% ▕████████████      ▏ 1.4 GB/2.0 GB   20 MB/s     29s

[링크 : https://ollama.com/library/llama3.2]

 

으아아 컴퓨터 녹는다 ㅋㅋㅋ

그래도 초당 2~3 단어 정도 나오는 느낌인데 이정도만 해도 그냥 대화하는 느낌 속도라

그거보다 빠르면 읽기도 전에 말이 더 빠르게 나올 듯?

 

Intel(R) Core(TM) i7-10510U CPU @ 1.80GHz

 

다시 실행하니 내가 받아둔 모델이 뜬다.

$ ollama  Select model to run: Type to filter...                                                                                                                                  Recommended                                                                          gemma4                                                                               Reasoning and code generation locally, ~16GB, (not downloaded)                   qwen3.5                                                                              Reasoning, coding, and visual understanding locally, ~11GB, (not downloaded)     kimi-k2.5:cloud                                                                      Multimodal reasoning with subagents                                              qwen3.5:cloud                                                                        Reasoning, coding, and agentic tool use with vision                              glm-5:cloud                                                                          Reasoning and code generation                                                    minimax-m2.7:cloud                                                                   Fast, efficient coding and real-world productivity                                                                                                                More                                                                               ▸ llama3.2                                                                                                                                                          ↑/↓ navigate • enter select • ← back

 

 

$ ollama --help Large language model runner Usage:   ollama [flags]   ollama [command] Available Commands:   serve       Start Ollama   create      Create a model   show        Show information for a model   run         Run a model   stop        Stop a running model   pull        Pull a model from a registry   push        Push a model to a registry   signin      Sign in to ollama.com   signout     Sign out from ollama.com   list        List models   ps          List running models   cp          Copy a model   rm          Remove a model   launch      Launch the Ollama menu or an integration   help        Help about any command Flags:   -h, --help         help for ollama       --nowordwrap   Don't wrap words to the next line automatically       --verbose      Show timings for response   -v, --version      Show version information Use "ollama [command] --help" for more information about a command.

 

단발성으로 질문하고 속도를 보려고 하는데

음.. eval rate가 실제 체감되는 속도인가? 12 token/s

이정도 수치는 되어야 많이 지루하지 않게 읽는 속도로 말이 생성되는 듯.

$ ollama run llama3.2:latest "안녕" --verbose 안녕하세요! (Hello!) I'm here to help you. How can I assist you today? total duration:       2.282485591s load duration:        143.825937ms prompt eval count:    27 token(s) prompt eval duration: 350.162008ms prompt eval rate:     77.11 tokens/s eval count:           21 token(s) eval duration:        1.73838491s eval rate:            12.08 tokens/s

 

gemma가 핫하긴 한데 용량이 커서 안돌것 같아서 3로 시도

[링크 : https://ollama.com/library/gemma3]

[링크 : https://ollama.com/library/gemma4]

 

제법 빠르고llama 보다는 괜찮은듯?

$ ollama run gemma3:1b --verbose >>> 안녕? 안녕하세요! 잘 지내세요? 😊 저는 오늘 당신의 도움이 필요하신 건가요? 궁금한  게 있으신가요? total duration:       1.695837878s load duration:        297.67234ms prompt eval count:    12 token(s) prompt eval duration: 164.822603ms prompt eval rate:     72.81 tokens/s eval count:           28 token(s) eval duration:        1.132524529s eval rate:            24.72 tokens/s

 

혹시나 gemma4를 16GB 메모리에서 시도해봤는데 메모리 부족으로 실패.

$ ollama run gemma4:e2b --verbose pulling manifest  pulling 4e30e2665218: 100% ▕██████████████████▏ 7.2 GB                          pulling 7339fa418c9a: 100% ▕██████████████████▏  11 KB                          pulling 56380ca2ab89: 100% ▕██████████████████▏   42 B                          pulling c6bc3775a3fa: 100% ▕██████████████████▏  473 B                          verifying sha256 digest  writing manifest  success  Error: 500 Internal Server Error: model requires more system memory (7.3 GiB) than is available (4.8 GiB)

 

프로세스 이것저것 죽이고 다시 시도하니 10token /s 정도 나오는 듯.

$ ollama run gemma4:e2b --verbose >>> 안녕 안녕하세요! 😊 total duration:       1.261211464s load duration:        302.800476ms prompt eval count:    17 token(s) prompt eval duration: 551.024817ms prompt eval rate:     30.85 tokens/s eval count:           4 token(s) eval duration:        226.924038ms eval rate:            17.63 tokens/s >>> 너에 대해서 소개해봐 Thinking... Thinking Process: 1.  **Analyze the Request:** The user asked "너에 대해서 소개해봐" (Tell  me about yourself). 2.  **Determine Identity & Core Facts (Self-Correction/Verification):**     *   Name: Gemma 4.     *   Developer: Google DeepMind.     *   Nature: Large Language Model (LLM).     *   Type: Open weights model.     *   Capabilities: Understands text/images, generates text.     *   Knowledge Cutoff: (Relevant for context, but usually stated  implicitly) 3.  **Structure the Introduction (Korean Context):** A good introduction  should cover:     *   Greeting.     *   Name/Identity.     *   Developer/Origin.     *   Nature/Function (What I do).     *   Capabilities (What I can process). 4.  **Draft the Response (Internal Monologue & Refinement):**     *   *Start warmly.* (안녕하세요!)     *   *State Identity.* (저는 Gemma 4입니다.)     *   *State Origin.* (Google DeepMind에서 개발했습니다.)     *   *State Nature.* (저는 대규모 언어 모델(LLM)입니다.)     *   *State Function/Role.* (사용자의 질문에 답하고, 정보를 제공하며, 다 양한 텍스트 기반 작업을 돕는 역할을 합니다.)     *   *State Modality/Input.* (텍스트와 이미지를 이해할 수 있습니다.)     *   *Concluding remark/Offer help.* (어떤 것을 도와드릴까요?) 5.  **Final Output Generation (Review against constraints):** The  generated response must adhere to the core identity (Gemma 4, LLM, Google  DeepMind, open weights). (The draft looks good and adheres to the required  identity.) ...done thinking. 안녕하세요! 저는 **Gemma 4**입니다. 저는 Google DeepMind에서 개발한 **대규모 언어 모델(LLM)**이며, 오픈 웨이트 (open weights) 모델로 공개되어 있습니다. **저의 주요 특징은 다음과 같습니다:** 1. **정체성:** 저는 방대한 데이터를 학습하여 인간의 언어를 이해하고 생성하며 , 다양한 질문에 답하고 정보를 제공하는 능력을 가지고 있습니다. 2. **기능:** 텍스트 기반의 대화, 요약, 번역, 작문, 코드 생성 등 다양한 언어  관련 작업을 수행할 수 있습니다. 3. **입력 능력:** 저는 텍스트와 이미지를 이해하고 처리할 수 있습니다. (다만 , 출력은 텍스트로만 생성합니다.) 4. **목표:** 사용자분들이 필요한 정보를 얻고, 창의적인 작업을 수행하며, 복잡 한 문제를 해결하는 데 도움을 드릴 수 있도록 설계되었습니다. 궁금한 점이 있거나 도움이 필요하시면 언제든지 말씀해주세요! 😊 total duration:       57.596689687s load duration:        297.105521ms prompt eval count:    35 token(s) prompt eval duration: 482.722245ms prompt eval rate:     72.51 tokens/s eval count:           614 token(s) eval duration:        56.261795129s eval rate:            10.91 tokens/s

 

+

intel 내장 그래픽(xe 나 arc 아님)은 openCL 가속 자체를 지원하지 않나?

$ clinfo Number of platforms                               0

[링크 : https://docs.ollama.com/gpu]

[링크 : https://dgpu-docs.intel.com/driver/client/overview.html#ubuntu-22.04]