You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thanks to zstd, the new package format is both smaller and decompresses faster. conda-forge channel has been releasing packages in .conda format for quite long.
To instruct conda-build to produce packages in .conda format, one can adding these lines to .condarc:
As two examples, using the command above, package size of the latest linux64 build pytorch-2.0.1-py3.11_cuda11.8_cudnn8.7.0_0 reduced from 1444MB to 992MB (-31%) and latest win64 build pytorch-2.0.1-py3.10_cuda11.8_cudnn8_0 reduced from 1395MB to 941MB (-33%).
Given that .conda format can significantly speed up installation (during both downloading and unpacking), and is already being widely used, I think it's worth switching to it?
The text was updated successfully, but these errors were encountered:
Conda 4.7 (released more than 4 years ago) introduced the new
.conda
package format, which uses zstd instead of bzip2 as compression algorithm.Thanks to zstd, the new package format is both smaller and decompresses faster.
conda-forge
channel has been releasing packages in.conda
format for quite long.To instruct
conda-build
to produce packages in.conda
format, one can adding these lines to.condarc
:or in command line:
And the
conda-package-handling
package providescph
command to convert packages from old.tar.bz2
format into.conda
format:As two examples, using the command above, package size of the latest linux64 build
pytorch-2.0.1-py3.11_cuda11.8_cudnn8.7.0_0
reduced from 1444MB to 992MB (-31%) and latest win64 buildpytorch-2.0.1-py3.10_cuda11.8_cudnn8_0
reduced from 1395MB to 941MB (-33%).Given that
.conda
format can significantly speed up installation (during both downloading and unpacking), and is already being widely used, I think it's worth switching to it?The text was updated successfully, but these errors were encountered: