[ACTION REQUIRED] Chainguard will disable CUDA Forward Compatibility in Triton Server images

TL;DR: Starting June 12th, 2026, Chainguard tritonserver and tritonserver-no-backend images will ship with CUDA Forward Compatibility disabled by default to reduce startup latency on hosts with modern NVIDIA drivers. If your hosts run older NVIDIA drivers (especially from a different major driver family than CUDA 13.x support), you must either upgrade your driver or explicitly opt back in by pointing LD_LIBRARY_PATH at the CUDA compatibility libraries directory in the container. Triton Server images geared towards a specific backend, like tritonserver-pytorch-backend or tritonserver-vllm-backend are unaffected by this change (as they already ship with forward compatibility disabled).

What is changing?

Today, Chainguard’s tritonserver and tritonserver-no-backend images are configured to use NVIDIA’s CUDA Forward Compatibility libraries by default. This is already the default behavior for our Triton Server images geared towards a specific backend, such as tritonserver-pytorch-backend and tritonserver-vllm-backend.

On March 6th, 2026, these Triton Server images will ship with Forward Compatibility disabled out-of-the-box.

CUDA Forward Compatibility is intended to allow applications built with a newer CUDA Toolkit to run on systems with an older NVIDIA GPU driver from a different major release family, by leveraging CUDA compatibility libraries (installed under a versioned path like /usr/local/cuda-<version>/compat/).

All other Chainguard images that provide CUDA, like PyTorch or NeMo, already ship with CUDA Forward Compatibility disabled by default.

Why is Chainguard making this change?

CUDA Forward Compatibility is valuable in environments with older drivers, but it can add overhead in environments where the host NVIDIA driver already meets the requirements for the CUDA runtime shipped in our images.

In particular, relying on PTX JIT compilation can introduce startup / first-run latency, because PTX may be compiled to native GPU code at runtime (often cached afterward).

Disabling automatic Forward Compatibility improves the default experience for customers running modern drivers, while still allowing customers on older drivers to opt back in.

How will this affect me?

If your host NVIDIA driver already supports CUDA 13.x (running nvidia-smi on the host reports a CUDA Version in the 13.x range): You do not need CUDA Forward Compatibility, and you may see improved startup performance by avoiding unnecessary compatibility/JIT paths.

 

If your host NVIDIA driver is older (running nvidia-smi on the host reports a CUDA Version < 13.0): Your deployment may fail to start unless you take action (either upgrade the driver or explicitly enable CUDA Forward Compatibility in your deployment).

What do I need to do?

Determine the version of CUDA supported by your host’s driver
On any host running Triton Server, connect to the host and run:

Look at the reported “CUDA Version” (this reflects the driver’s maximum supported version of CUDA).

If the reported CUDA version is < 13.0, you have two options:

Option A (recommended): Upgrade the NVIDIA driver on the host (if possible)
Upgrade to a driver that supports CUDA 13.x-level workloads so Forward Compatibility is not needed.

Option B: Opt back in to CUDA Forward Compatibility
Set LD_LIBRARY_PATH in the container’s environment to include the directory containing compatibility shipped in the image (examples shown below):

Kubernetes manifest:

Dockerfile:

FAQs

Which images are impacted?

Only our tritonserver and tritonserver-no-backend images. All other Triton Server images provided by Chainguard already ship with CUDA Forward Compatibility disabled to provide the best possible out-of-the-box experience.

What is CUDA Forward Compatibility?

A mechanism used to run applications built on newer versions of CUDA on systems with older NVIDIA GPU drivers from different major release families.

Does CUDA Forward Compatibility work with every NVIDIA GPU?

No. NVIDIA documents CUDA Forward Compatibility as applicable to NVIDIA Data Center GPUs, select NGC Server Ready RTX SKUs, and Jetson boards.

If I still need CUDA Forward Compatibility, can I enable it?

Yes. Set LD_LIBRARY_PATH in the container’s environment to include the directory containing compatibility libraries shipped in the image.

Need help or have questions?

We're here to help - visit our support portal at support.chainguard.dev.

- Chainguard Team