git-lfs-faq - Man Page

No, it doesn’t.  Unfortunately, Git itself doesn’t specify a way to make .gitattributes patterns apply to files of a certain size and we rely on the .gitattributes file to specify which files are tracked by Git LFS.

You can use the --above option to git lfs migrate import to migrate all files that at the specified time are larger than a certain size.  However, if your files change to be smaller or larger in the future, or you add more files in the future that are larger than the limit you specified, you will have to track them manually.

For these reasons, we recommend using patterns rather than --above.

Git LFS itself handles these files just fine.  However, Git LFS is usually invoked by Git, and until Git 2.34, Git itself on Windows didn’t handle files using smudge and clean filters (like Git LFS) that are larger than 4 GiB. So you can update Git for Windows to 2.34 to natively support these file sizes.

On older versions, set GIT_LFS_SKIP_SMUDGE to 1 and run git lfs pull to pull down the LFS files.  This bypasses Git’s smudging functionality and therefore avoids its limitations.

Git LFS stores small text files called pointer files in the repository instead of your large files, which it stores elsewhere.  These pointer files usually start with the line version https://git-lfs.github.com/spec/v1.

Normally, if you’ve run git lfs install at least once for your user account on the system, then Git LFS will be automatically invoked by Git when you check out files or clone a repository and this won’t happen.  However, if you haven’t, or you’ve explicitly chosen to skip that behaviour by using the --skip-smudge option of git lfs install, then you may need to use git lfs pull to replace the pointer files in your working tree with large files.

This can happen if someone made a commit to a file that’s tracked by Git LFS but didn’t have Git LFS properly set up on their system.  The objects that were checked into the repository are Git objects, not the pointers to Git LFS objects, and when Git checks these files out, it shows them as modified.

There are also several other possible ways to encounter this problem, such as an incomplete migration of your repository.  For example, you should not use git lfs track to track patterns that are already in your repository without running git add --renormalize ., since that can lead to this problem.

Users frequently find that this cannot be changed by doing git reset --hard or other techniques because Git then checks the files out and marks them as modified again.  The best way to solve this problem is by fixing the files and the committing the change, which you can do with the following on an otherwise clean tree:

$ git add --renormalize .
$ git commit -m "Fix broken LFS files"

This requires that every branch you want to fix have this done to it.

To prevent this from reoccurring in the future, make sure that everyone working with large files on a project has run git lfs install at least once.  The command git lfs fsck --pointers BASE..HEAD (with suitable values of BASE and HEAD) may be used in your CI system to verify that nobody is introducing such problems.

If you want to track files that already exist in a repository, you need to do two things.  First, you need to use git lfs track (or a manual modification of .gitattributes) to mark the files as LFS files.  Then, you need to run git add --renormalize . and commit the changes to the repository.

If you skip this second step, then you’ll end up with files that are marked as LFS files but are stored as Git files, which can lead to files which are always modified, as outlined in the FAQ entry above.  Note also that this doesn’t change large files in your history.  To do that, use git lfs migrate import --everything instead, as specified in one of the entries below.

You can run git config diff.lfs.textconv cat, which will produce normal diffs if your files are text files.

If the above solution is too broad, each entry in the .gitattributes file can be customized by creating a custom global converter:

$ git config --global diff.lfstext.textconv cat

Any given .gitattributes entry for large text files can be customized to use this global text converter (e.g., patch files), whereas binary formats can continue to use the conventional lfs diff tool, like so:

$ cat .gitattributes
....
*.bin filter=lfs diff=lfs merge=lfs -text
*.patch filter=lfs diff=lfstext merge=lfs -text
....

Be advised that all developers sharing this repo with such a modified .gitattributes file must have similarly configured the lfstext text converter, whether globally or on a per repository basis.

If you’d like to stop using Git LFS and switch back to storing your large files in the plain Git format, you can do so with git lfs migrate export --everything.  Note that you will need to provide an appropriate --include option to match all the patterns that you currently have tracked in any ref.

This also rewrites history, so the Git object IDs of many, if not all, of your objects will change.

GitHub rejects large files anywhere in the history of your repository, not just in the latest commit.  If you’re still seeing this message, then you have some large files somewhere in your history, even if in the latest commits you’ve moved them to Git LFS.

To fix this, you can use git lfs migrate import --everything with an appropriate --include argument.  For example, if you wanted to move your .jpg and .png files into Git LFS, you can do that with git lfs migrate import --everything --include="*.jpg,*.png".  More complicated patterns are possible: run git help gitattributes for more information on valid patterns.  Note that if you’re specifying directories, using slashes is mandatory: backslashes are not allowed as path separators.

Recent versions of Jenkins have set core.hooksPath to various values, notably NUL: on Windows, with the goal of disabling hooks.  This is not a valid path on Windows, nor a valid value for this configuration option, so when git lfs install runs and Git LFS attempts to install hooks, the operation fails.

The easiest way to solve this problem is by using the --skip-repo option to git lfs install, which skips the installation of the hooks.  Despite the name, it can be successfully combined with --local if you need that option.

Note that this prevents things like git push from pushing LFS objects and locked files from being read only, since those are implemented by hooks.  If you need that functionality, you should review the Jenkins documentation about how to properly configure the environment in such a situation so that hooks can be used.

When you run git archive with only a subdirectory, such as git archive HEAD:src, Git resolves the revision (in this case, HEAD:src) to a tree, and only processes items in that tree.  Because the .gitattributes file is typically only in the root of the repository, Git doesn’t even see that file, which controls whether files are considered LFS files, and hence doesn’t consider any of the files in the directory as LFS files, and thus doesn’t invoke Git LFS at all.

Since Git LFS doesn’t even get invoked in this case, there’s no way to change how this works.  If you just want to include the single subdirectory without stripping the prefix, you can do this: git archive -o archive.tar.gz --prefix=archive/ HEAD src.  If you do want to strip the subdirectory name (src) in this case, one option if you have the libarchive tar (available on Windows and macOS as tar, and usually on Linux as bsdtar) is to do something like this script:

#!/bin/sh

# With trailing slash.
ARCHIVE_PREFIX="archive/"
# Without trailing slash.
SOURCE_PREFIX="src"
# Without directory or file components.
REVISION="HEAD"

temp=$(mktemp -d)

git archive --prefix="$ARCHIVE_PREFIX" "$REVISION" "$SOURCE_PREFIX" | bsdtar -C "$temp" -xf -
bsdtar -s "!^\./!$ARCHIVE_PREFIX!" --format=pax -czf archive.tar.gz -C "$temp/$ARCHIVE_PREFIX$SOURCE_PREFIX" .
rm -fr "$temp"