mamba: critical libmamba Package cache error. When building with micromamba

When building the following environment the mamba installation works fine, but the micromamba installation gives an error:

name: test
channels:
  - default
  - pytorch
  - nvidia
  - conda-forge
dependencies:
  - python=3
  - pytorch 
  - torchvision 
  - torchaudio 
  - cudatoolkit=11.7
  - numpy
  - opencv
  - pillow
  - matplotlib
  - scipy
  - scikit-learn
  - pip
  - tensorboard
  - pip:
    - 'git+https://github.com/facebookresearch/detectron2.git'
    - natsort
    - torch-tb-profiler
    - imagesize

The error in question is as follows (includes all warning, please let me know if something else is needed):

warning  libmamba [cuda-cuobjdump-11.7.91-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-cuxxfilt-11.7.91-0
warning  libmamba [cuda-cuxxfilt-11.7.91-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-nvcc-11.7.99-0
warning  libmamba [cuda-nvcc-11.7.99-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-nvprune-11.7.91-0
warning  libmamba [cuda-nvprune-11.7.91-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-cccl-11.7.91-0
Linking cuda-driver-dev-11.7.99-0
Linking cuda-cupti-11.7.101-0
warning  libmamba [cuda-cupti-11.7.101-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-nvtx-11.7.91-0
warning  libmamba [cuda-nvtx-11.7.91-0] The following files were already present in the environment:
    - LICENSE
Linking cuda-nsight-12.0.78-0
warning  libmamba [cuda-nsight-12.0.78-0] The following files were already present in the environment:
    - LICENSE
Linking nsight-compute-2022.4.0.15-0
warning  libmamba [nsight-compute-2022.4.0.15-0] The following files were already present in the environment:
    - LICENSE
Linking libcufile-1.5.0.59-0
warning  libmamba Invalid package cache, could not read 'paths.json' from '/home/stefan/micromamba/pkgs/libcufile-1.5.0.59-0'
    
error    libmamba Cannot find a valid extracted directory cache for 'libcufile-1.5.0.59-0.tar.bz2'
critical libmamba Package cache error.

Versions:

$mamba --version
mamba 1.0.0
conda 22.9.0
$micromamba -v
Version: 1.1.0

First encountered this problem when trying to build a docker container using the micromamba docker, but the error was present in version 1.0.0 and outside the docker environment as well.

Tested on Ubuntu 22.04

About this issue

  • Original URL
  • State: closed
  • Created 2 years ago
  • Comments: 19 (6 by maintainers)

Commits related to this issue

Most upvoted comments

Adding libcufile=1.4.0.31 to my dependency list in my conda.yaml worked for me

I had the same issue. Until this is fixed, I am using a workaround using a mamba binary installed through micromamba … ugh 😉 . While not recommended, at least this allows for faster Dockerfile building. Thanks @urucoder, for the hint!

Same issue here. Would be great if you could fix that. We downgraded the lib. This works to resolve the dependencies, but no idea what the impact will be.

  • nvidia::libcufile>=1.4,<1.5

I can confirm suffering the same issue on my side, using the mamba binary from the micromamba installation works like a charm.

Failing

micromamba install pytorch::pytorch=1.13.1=py3.9_cuda11.7_cudnn8.5.0_0 -c pytorch -c nvidia

Installing successfully

micromamba create -n env conda-forge::mamba /opt/micromamba/envs/env/condabin/mamba install pytorch::pytorch=1.13.1=py3.9_cuda11.7_cudnn8.5.0_0 -c pytorch -c nvidia