Tensorflow-gpu doesnt install properly "bad address"

[jsmith432]

Trying to install in a QNAP NAS (container station), Docker Compose example, with a NVIDIA video card. But the library is not being set up correctly. How do I fix this? Do you know if I need to mount something else?

init: tensorflow-gpu
/scripts/install-tensorflow.sh gpu
Installing TensorFlow 2.18.0 for AMD64-GPU in "/usr"...
Downloading amd64 libs from "https://dl.photoprism.app/tensorflow/amd64/libtensorflow-amd64-gpu-2.18.0.tar.gz". Please wait.
Extracting "/tmp/amd64/libtensorflow-amd64-gpu-2.18.0.tar.gz" to "/usr".
tar: include/tensorflow/c: Cannot change mode to rwxr-xr-x: Bad address
tar: include/tensorflow/core/platform: Cannot change mode to rwxr-xr-x: Bad address
tar: include/tensorflow/core: Cannot change mode to rwxr-xr-x: Bad address
tar: include/tensorflow: Cannot change mode to rwxr-xr-x: Bad address
tar: include/tsl/platform: Cannot change mode to rwxr-xr-x: Bad address
tar: include/tsl: Cannot change mode to rwxr-xr-x: Bad address
tar: include/xla/tsl/c: Cannot change mode to rwxr-xr-x: Bad address
tar: include/xla/tsl: Cannot change mode to rwxr-xr-x: Bad address
tar: include/xla: Cannot change mode to rwxr-xr-x: Bad address
tar: lib/libtensorflow.so: Cannot change mode to r-xr-xr-x: Bad address
tar: lib/libtensorflow.so.2: Cannot change mode to r-xr-xr-x: Bad address
tar: lib/libtensorflow_framework.so: Cannot change mode to r-xr-xr-x: Bad address
tar: lib/libtensorflow_framework.so.2: Cannot change mode to r-xr-xr-x: Bad address
tar: include/tensorflow/c/eager: Cannot change mode to rwxr-xr-x: Bad address
tar: Exiting with failure status due to previous errors
make: *** [Makefile:38: install-tensorflow-gpu] Error 2

I am also getting other errors:
Updating permissions of keys in '/etc/ssl/private'...
chmod: changing permissions of '/etc/ssl/private/photoprism.issuer.key': Bad address
chmod: changing permissions of '/etc/ssl/private/photoprism.key': Bad address
chmod: changing permissions of '/etc/ssl/private/photoprism.pfx': Bad address

[omerdduran]

Hello. The error you encountered on your QNAP NAS may be caused by file system compatibility issues while PhotoPrism attempts to extract TensorFlow libraries and modify file permissions. To resolve this issue, you should add PHOTOPRISM_DISABLE_CHOWN: “true” to the environment variables in your docker-compose.yml file. This disables automatic permission updates via chmod and chown, which trigger these errors.

[jsmith432]

Thanks for the reply @omerdduran . Changing PHOTOPRISM_DISABLE_CHOWN: “true”, has no effect. Same errors happen.
Seems to me that the script copies the files to /usr, and that is not mounted to the NAS, so I assumed that is not part of the storage that is persisted on the NAS. Am I incorrect to assume so?

[lastzero]

Those "bad address" errors occur when the container attempts to (1) extract TensorFlow libraries into /usr and (2) change the permissions of keys under /etc/ssl/private. In some QNAP Container Station setups, the root file system/overlay does not tolerate these runtime writes. As a workaround, set:

PHOTOPRISM_INIT: ""
PHOTOPRISM_DISABLE_TLS: "true"
PHOTOPRISM_DISABLE_CHOWN: "true"

[jsmith432]

Thanks @lastzero . I was trying to use the gpu to help with the indexing. I assume that when I blank the init, those drivers don't get installed and I dont use that. Does that mean that in my current configuration I can't use that unless it is somehow bundled in the image?

[lastzero]

It's a mechanism integrated with your QNAP NAS, probably for security purposes, and I can't tell you why it's set up this way. It's not a bug in our software. One possible workaround is to build a custom image, but that might be too much effort.

[jsmith432]

Strangenly enough I can just get into the container and do a chmod 755 to files and the links seems to be already set with the right permission.

lrwxrwxrwx 1 root root 18 Dec 31 1999 libtensorflow.so -> libtensorflow.so.2
lrwxrwxrwx 1 root root 23 Dec 31 1999 libtensorflow.so.2 -> libtensorflow.so.2.18.0
-rwxr-xr-x 1 root root 919168672 Dec 31 1999 libtensorflow.so.2.18.0

root@250707:/usr/lib$ chmod 755 libtensorflow_framework.so.2.18.0
root@250707:/usr/lib$ ls -la libtensorflow_framework.so.2.18.0
-rwxr-xr-x 1 root root 152341992 Dec 31 1999 libtensorflow_framework.so.2.18.0

Are they links that get error out? Seems to me that it extracts, setting the permission with --mode=755 hardcoded on the install script and just runs ldconfig at the end.