Cannot use fractional g4 instance types on Google Batch

Google Cloud Compute Infrastructure
1

I am trying to use the fractional RTX PRO 6000 instance types (g4-standard-6, g4-standard-12, g4-standard-24) on Google Batch.

When creating a job on the Cloud Console, the fractional GPUs can be selected, but the instance type cannot. A g4-standard-48 instance is used:

Screenshot 2026-04-10 at 10.26.11 AM
Screenshot 2026-04-10 at 10.26.11 AM1222×734 59 KB

When submitting a job via the Batch API, the accelerator cannot be attached. If an accelerator is not specified, the instance can be used, but GPU appears not to be available:

400 machine_type field is invalid. machine type g4-standard-12 is not compatible with accelerators [type:“nvidia-rtx-pro-6000” count:1] error: generic::invalid_argument: Accelerator field is invalid. Machine type g4-standard-12 does not support accelerator with type nvidia-rtx-pro-6000 and GPU count 1. Please make sure that the configuration meets this requirement: GPU machine types  |  Compute Engine  |  Google Cloud Documentation.

2

Hello timpalpant,

Thanks for reporting this. I have confirmed that this is a known issue on our end affecting g4-standard-6, g4-standard-12, and g4-standard-24.

As a workaround, please try using g4-standard-48 (or higher) for your current jobs. We are working on a fix for these specific instance types.

Please let me know if you face any further issues with this workaround.

3

Hello Timpalpant,

Fractional G4 GPUs are now available and can be configured via both the UI and gcloud.

If you use gcloud, please set the GPU count to 1 rather than specifying 1/2, 1/4, or 1/8. This will automatically provision a VM with the fractional GPU, which you can then verify in the VM configuration.

Please let me know if you need any help with this.

4

Hi @Giorgi_Botkoveli,

Thanks for the update and support to quickly add this feature. It is working for me.

Note: You must install and use the NVIDIA driver for fractional instances, as described here: Install GPU drivers  |  Compute Engine  |  Google Cloud Documentation

5

Hello Timpalpant,

As a workaround for GPU driver installations, Create and run a job that uses GPUs  |  Batch  |  Google Cloud Documentation might be useful.

6

Thanks @Giorgi_Botkoveli, it is working well for us!

We do observe one strange issue on fractional G4 instances: after 24h of runtime, our jobs failed with CUDA_ERROR_DEVICE_NOT_LICENSED (102). Do you happen to know if this is a known limitation of fractional instances, or if there is something we need to do differently?