Nvidia驱动程序错误-CUDA驱动程序版本不足以支持CUDA运行时版本

时间:2020-01-02 14:47:50

标签: cuda nvidia-docker

我正试图称呼Laia-HRW的深度学习工具包:https://github.com/jpuigcerver/Laia

这是我的代码:

INPUT_DIR=`pwd`/RecognitionHand/dir_input
OUTPUT_DIR=`pwd`/RecognitionHand/dir_output
CHAR_TRANSCRIBE_FILE=char.txt
WORD_TRANSCRIBE_FILE=word.txt

rm $INPUT_DIR/filelist/filenames.lst

ls -d -1 $INPUT_DIR/images/* > $INPUT_DIR/filelist/filenames.lst

COMMAND="decode --batch_size 20  --log_level info   --symbols_table \
    $INPUT_DIR/symbtable/symbs.txt \
    $INPUT_DIR/model/model_htr.t7 \
    $INPUT_DIR/filelist/filenames.lst> $OUTPUT_DIR/$CHAR_TRANSCRIBE_FILE";


# local volumes mapped to the docker volumes
OPTS=( -u $(id -u):$(id -g) );
[ -d "/home" ]  && OPTS+=( -v /home:/home );
[ -d "/mnt" ]   && OPTS+=( -v /mnt:/mnt );
[ -d "/media" ] && OPTS+=( -v /media:/media );
[ -d "/tmp" ]   && OPTS+=( -v /tmp:/tmp );


# call the GPU docker for transcribing
docker run --rm -t "${OPTS[@]}" laia:active \
  bash -c "cd $(pwd) && PATH=\" .:$PATH:\$PATH\" laia-$COMMAND";

最后一个docker命令引用了nvidia-docker,我遇到了这个奇怪的错误:

THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-918/cutorch/lib/THC/THCGeneral.c line=66 error=35 

: CUDA driver version is insufficient for CUDA runtime version
[2020-01-02 14:43:45  WARN] /opt/torch/share/lua/5.1/laia/util/base.lua:39: Optional lua module "cutorch" was not found!
[2020-01-02 14:43:45  WARN] /opt/torch/share/lua/5.1/laia/util/base.lua:39: Optional lua module "cunn" was not found!
[2020-01-02 14:43:45  WARN] /opt/torch/share/lua/5.1/laia/util/base.lua:39: Optional lua module "laia.util.cudnn" was not found!
[2020-01-02 14:43:45  WARN] /opt/torch/share/lua/5.1/laia/util/base.lua:39: Optional lua module "laia.ImageDistorter" was not found!
/opt/torch/bin/luajit: /opt/torch/lib/luarocks/rocks/laia/scm-1/bin/laia-decode:16: attempt to call field 'registerOptions' (a nil value)
stack traceback:
    /opt/torch/lib/luarocks/rocks/laia/scm-1/bin/laia-decode:16: in main chunk
    [C]: at 0x00405d50

为什么会这样?请问有人在运行nvidia-docker时遇到了类似的错误吗?

1 个答案:

答案 0 :(得分:1)

CUDA驱动程序版本不足于CUDA运行时版本

表示系统的nvidia驱动程序与您下载的docker映像内的运行时不兼容。您必须(至少)匹配这些版本。 另一个重要的事情是检查您要使用的工具所必需的cuda运行时。假设您遵循必需的版本并下载了正确的Docker映像,则需要更新系统nvidia驱动程序以匹配Docker映像。