gcsfuse not working with custom container in Vertex AI on NVIDIA_L4

[ronaldpanape]

System (please complete the following information):

Vertex AI on NVIDIA_L4

Dockerfile setup

# Set up the GCSFuse repository
#RUN export GCSFUSE_REPO=gcsfuse-`lsb_release -c -s` && \
#    echo "deb https://packages.cloud.google.com/apt $GCSFUSE_REPO main" | tee /etc/apt/sources.list.d/gcsfuse.list && \
#    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key add -

# Install GCSFuse
#RUN apt-get update && \
#    apt-get install -y fuse gcsfuse

Error:

{"timestamp":{"seconds":1706802567,"nanos":766028008},"severity":"TRACE","message":"fuse_debug: Beginning the mounting kickoff process"}
{"timestamp":{"seconds":1706802567,"nanos":766076620},"severity":"TRACE","message":"fuse_debug: Parsing fuse file descriptor"}
{"timestamp":{"seconds":1706802567,"nanos":766087548},"severity":"TRACE","message":"fuse_debug: Preparing for direct mounting"}
{"timestamp":{"seconds":1706802567,"nanos":766127226},"severity":"TRACE","message":"fuse_debug: Directmount failed. Trying fallback."}
{"timestamp":{"seconds":1706802567,"nanos":766437137},"severity":"TRACE","message":"fuse_debug: Creating a socket pair"}
{"timestamp":{"seconds":1706802567,"nanos":766535546},"severity":"TRACE","message":"fuse_debug: Creating files to wrap the sockets"}
{"timestamp":{"seconds":1706802567,"nanos":766552608},"severity":"TRACE","message":"fuse_debug: Starting fusermount/os mount"}
/usr/bin/fusermount: fuse device not found, try 'modprobe fuse' first
{"timestamp":{"seconds":1706802567,"nanos":768124241},"severity":"ERROR","message":"Error while mounting gcsfuse: mountWithStorageHandle: Mount: mount: running /usr/bin/fusermount: exit status 1\n"}
mountWithArgs: mountWithStorageHandle: Mount: mount: running /usr/bin/fusermount: exit status 1

[Tulsi Shah]

Hi @ronaldpanape,

The "fuse device not found" error usually means the FUSE kernel module isn't loaded. Here's how to fix that:

1. Install FUSE:
   Ubuntu/Debian: sudo apt-get install fuse
   CentOS/Fedora/RHEL: sudo yum install fuse
2. Load the module: sudo modprobe fuse
3. run container with --privilege mode.

Let me know if this resolves the issue. If not, please provide more details about your system and the command you were trying to run.

Thanks,
Tulsi Shah

[ronaldpanape]

Hi @Tulsishah

I am using Vertex AI, as seen below how do i get that working in that environment

    worker_pool_specs=[
        {
            "container_spec": {
                "image_uri": f"us-docker.pkg.dev/{project_id}/training/axolotl:v0.4.0",
                "command": ["/bin/bash", "-c"],
                "args": [full_command],
            },
            "replica_count": 1,
            "machine_spec": {
                "machine_type": machine_type,
                "accelerator_type": accelerator_type,
                "accelerator_count": accelerator_count,
            },
        }
    ],

[Tulsi Shah]

Hi @ronaldpanape,

I think you can add following in docker file and test

RUN apt-get install fuse && \
                modprobe fuse 

This issue seems like it might be related to Vertex AI, as gcsfuse require to install fuse first and fuse itself is not working here. If the above solution doesn't resolve the problem, I recommend raising an issue directly on the Vertex AI issue tracker for further investigation.

Thanks,
Tulsi Shah

[Nitin Garg]

@ronaldpanape this looks like an issue in the VertexAI training VM, so I checked with the VertexAI training team.

Copying Vertex AI team's response from internal portal:

Customers don't need to install GCSFuse themselves. As a managed training solution, Vertex AI Training installs and mounts GCSFuse for customers. Customers can directly read GCS in their containers. See this documentation for instructions: https://cloud.google.com/vertex-ai/docs/training/cloud-storage-file-system.

If the customers run into any further issues, please share with us the project number and job ID, our oncalls are more than happy to help.

I'll try to find out if the vertex AI team can directly respond to the issue somehow.

[Nitin Garg]

@ronaldpanape the vertexAI team suggested that you use gcsfuse in vertexAI training using the documentation here and open a support ticket there, if it does not work, and then vertex AI team can actively support you.

[ronaldpanape]

this works for me

[Miguel Angel Alba Acosta]

@gargnitingoogle is there any way users can configure in vertex AI the cache options from GCSFuse as stated here?, I am having data starvation issues on the multi-node distributed training setting in vertex AI pipelines, i.e Low GPU utilization in some nodes

I wonder if GCSFuse can be installed in the container, the bucket is mounted and set an entrypoint with the cache file, stat and type options before the training runs the command in vertex AI

[Nitin Garg]

@miguelalba96 the file cache feature that you pointed to, was added in gcsfuse v2.0.0, which is not used by the prebuilt images provided by Vertex AI yet (but it will be in a future version soon). So, if you are using managed vertex AI, then there is no way to do it as of today.

is there any way users can configure in vertex AI the cache options from GCSFuse?

It depends on how gcsfuse is installed and configured in your setup. If you have control over those two parameters, then it might work out, which is not possible for vertex AI managed jobs.

I wonder if GCSFuse can be installed in the container ...

I don't think it is possible in a managed vertex AI job and you might have to create your own custom container. When you install and configure GCSFuse yourself in a custom container, then remember to install gcsfuse v2.0.0 (instructions) and remember that only a privileged container is supported for making GCSFuse work (documented in troubleshooting).

[Nitin Garg]

@miguelalba96 as you've opened a new issue #1830 for the same question, let's continue this discussion there.