Tensorflow: CUDA driver version is insufficient for CUDA runtime. Can I install drivers in container?

[Robert Lugg]

I'm getting the error when running a newer version of TensorFlow.  I suspect the drivers on my host computer are too old for this CUDA library.

Is there any way to leave the host drivers as-is and (somehow) upgrade them within the container?

Regards.

[Kandes, Martin]

Hi Robert,

Unfortunately, no. You're limited by the the drivers on the host system.

Marty

---

From: Robert Lugg <rober...@gmail.com>
Sent: Thursday, March 14, 2019 1:42:07 PM
To: singularity
Subject: [Singularity] Tensorflow: CUDA driver version is insufficient for CUDA runtime. Can I install drivers in container?

 

I'm getting the error when running a newer version of TensorFlow.  I suspect the drivers on my host computer are too old for this CUDA library.

Is there any way to leave the host drivers as-is and (somehow) upgrade them within the container?

Regards.

--
You received this message because you are subscribed to the Google Groups "singularity" group.
To unsubscribe from this group and stop receiving emails from it, send an email to singularity...@lbl.gov.

[Gregory M. Kurtzer]

Are you using the `--nv` runtime Singularity option?

--

You received this message because you are subscribed to the Google Groups "singularity" group.
To unsubscribe from this group and stop receiving emails from it, send an email to singularity...@lbl.gov.

--

Gregory M. Kurtzer

CEO, Sylabs Inc.

[David Godlove]

Installing an nvidia driver really does 2 things.  It installs a kernel module, and it copies a bunch of user-space libraries onto the host.  Your container shares the kernel (and thus the nvidia kernel module) with the host system.  The user-space libraries inside the container have to match the kernel module version exactly.  So you really can't upgrade the driver in the container without also updating the driver on the host system.  Sorry.