CUDA

CUDA

Driver Update

$ sudo ubuntu-drivers list --gpgpu
$ sudo ubuntu-drivers install --gpgpu nvidia:550-server
$ sudo reboot

재부팅 후 nvidia-smi가 또 안되는 오류가 발생하여 가이드에 따라 기존 드라이버 모두 언인스톨 후 cuda만 실행했다.

CUDA Driver Update

CUDA Toolkit 12.4 (Apr 2024)
https://developer.nvidia.com/cuda-toolkit-archive runfile(local) 방식이 가장 간단

$ sudo apt install build-essential
$ wget https://developer.download.nvidia.com/compute/cuda/12.4.1/local_installers/cuda_12.4.1_550.54.15_linux.run
$ chmod +x cuda_12.4.1_550.54.15_linux.run
$ sudo ./cuda_12.4.1_550.54.15_linux.run

CUDA Installer가 drvier까지 같이 설치해주는 것으로 보이는데 확인 필요

언인스톨은 다음과 같다.

# To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-12.4/bin
# To uninstall the NVIDIA Driver, run nvidia-uninstall
$ sudo apt-get purge nvidia*
$ sudo apt-get autoremove

$ sudo systemctl isolate multi-user.target
$ sudo modprobe -r nvidia-drm

$ sudo reboot

nvidia-container(Docker) 설치

apt에서 hash 에러가 발생하므로 apt 주소 변경:

$ sudo vi /etc/apt/sources.list
:%s/kr.archive.ubuntu.com/mirror.kakao.com/
$ sudo apt update

Install Docker with NVIDIA support 참고

트러블슈팅

CUDA를 설치하면 nvidia driver 버전은 맞춰서 함께 올라간다.

nvidia-smi실행시,

NVML: Driver/library version mismatch

갑자기 드라이버 오류 발생, 리부팅으로 해결1

Dockerfile

CUDA 기반 Dockerfile:

FROM nvcr.io/nvidia/cuda:12.4.0-devel-ubuntu22.04

# Set the Timezone to KST
ENV TZ=Asia/Seoul

# Avoid question during docker build.
ARG DEBIAN_FRONTEND=noninteractive

# Change apt repo to kakao due to a hash error.
RUN sed -i "s/archive.ubuntu.com/mirror.kakao.com/g" /etc/apt/sources.list

# Install additonal packages.
RUN apt-get update && \
    apt-get install -y --no-install-recommends locales locales-all tzdata && \
    locale-gen en_US en_US.UTF-8 && update-locale LC_ALL=en_US.UTF-8 LANG=en_US.UTF-8 && \
    locale && \
    apt-get install -y --no-install-recommends \
          cmake \
          vim \
          silversearcher-ag \
          file \
          screen \
          wget \
          git \
          python3-dev \
          python3-pip \
          tree && \
    rm -rf /var/lib/apt/lists/* && \
    apt-get clean

# Default `python` runs `python3`
RUN echo 'alias python=python3' >> ~/.bashrc

# Default workdir is `/home/xxxx` on the host.
WORKDIR /xxxx

llama.cpp 빌드:

$ cmake .. -DLLAMA_CUDA=on -DLLAMA_CUDA_F16=1 -DCMAKE_CUDA_ARCHITECTURES=89
$ cmake --build . --config Release --parallel 8

huggingface-cli(HF_TOKEN 필요):

$ pip install -U "huggingface_hub[cli]"
$ huggingface-cli download google/gemma-2b-it

BLAS

cuBLAS가 가장 빠르다. 슈트라센 알고리즘으로 \(O(n^{2.37188})\)으로 가능함에도 불구하고.

Last Modified: 2024/04/22 14:43:55

is a collection of Papers I have written.
© 2000 - Sang-Kil Park Except where otherwise noted, content on this site is licensed under a CC BY 4.0.
This site design was brought from Distill.