Terra GPU environments currently install NVIDIA driver 535.154.05, as configured in gce-init.sh.
We are running an inference stack using vLLM and PyTorch inside a custom container on Terra T4 GPUs. The container supplies the CUDA user-space libraries, but it must remain compatible with the NVIDIA driver installed on the Terra host.
The current R535 driver supports our CUDA 12 configuration, but it prevents us from qualifying CUDA 13 builds, which require R580 or newer. It also requires us to maintain a Terra-specific runtime combination as vLLM and PyTorch adopt newer CUDA versions.
Is there a plan to update the pinned driver and COS image to R580 or newer? Are there plans to manage the driver through a more regularly updated configuration instead of pinning it in the VM initialization script? Would users need to recreate their cloud environments to receive an updated driver?
Thanks for looking into this!
Terra GPU environments currently install NVIDIA driver
535.154.05, as configured ingce-init.sh.We are running an inference stack using vLLM and PyTorch inside a custom container on Terra T4 GPUs. The container supplies the CUDA user-space libraries, but it must remain compatible with the NVIDIA driver installed on the Terra host.
The current R535 driver supports our CUDA 12 configuration, but it prevents us from qualifying CUDA 13 builds, which require R580 or newer. It also requires us to maintain a Terra-specific runtime combination as vLLM and PyTorch adopt newer CUDA versions.
Is there a plan to update the pinned driver and COS image to R580 or newer? Are there plans to manage the driver through a more regularly updated configuration instead of pinning it in the VM initialization script? Would users need to recreate their cloud environments to receive an updated driver?
Thanks for looking into this!