Runpod rebuilds PyTorch templates to cover versions 2.6 through 2.13
Runpod has rebuilt its PyTorch templates, which now cover versions 2.6 through 2.13 across multiple CUDA and Ubuntu versions. The GPU cloud provider also added a CUDA dropdown to the Pod deploy flow so users can match a template to their GPU and workload.
- Runpod's PyTorch templates now cover versions 2.6 through 2.13 across multiple CUDA and Ubuntu versions.
- A new CUDA dropdown in the Pod deploy flow lets users match a template to their GPU and workload.
- Global Volumes mount like network volumes but run on object storage, according to Runpod.
- A Runpod guide explains why Serverless balances can drop even when jobs fail.
- The source text gives no details on the OpenCV partnership named in its title.
PyTorch templates span 2.6 to 2.13
According to Runpod, its PyTorch templates have been rebuilt and now span PyTorch 2.6 through 2.13. The coverage extends across multiple CUDA and Ubuntu versions, so a single PyTorch release can be paired with different software stacks depending on what a project needs.
To make that choice easier, the Pod deploy flow now includes a CUDA dropdown. Runpod says the purpose is to let users match a template to their GPU and workload, rather than accepting a fixed default. For teams that depend on a specific PyTorch version or need compatibility with a particular CUDA release, this reduces the need to adjust an environment after a Pod has started.
The source text does not list which CUDA or Ubuntu versions are included, nor does it say how existing templates are affected, so users should check the Runpod console for the exact options.
Global Volumes run on object storage
The same Runpod blog material also addresses Global Volumes. Runpod stresses that they are not "network volumes, but better." Instead, the company describes them as something with a different purpose.
Global Volumes mount like a network volume, but they run on object storage. According to Runpod, that difference changes what users should put on them. The source text does not go into which kinds of data suit each option, but the message is clear: although the mounting experience looks familiar, the underlying storage type should guide what is stored there.
Serverless billing when jobs fail
A third item covers how Runpod Serverless billing works when a job fails to complete. The description promises to explain why an account balance can drop even when nothing succeeds, how to find where the money went, and how to keep it from happening again.
The text provided gives no figures, rates or specific rules for failed jobs, so the details sit in the full guide. The topic is relevant for anyone running Serverless workloads who has seen charges without successful results.
What it means for Runpod users
For developers deploying Pods, the practical change is more control over the software environment at launch. A wider range of PyTorch versions and a CUDA dropdown should help match environments to hardware and project requirements. The Global Volumes and billing explainers are aimed at helping users make informed decisions about storage and costs.
The page is titled "Runpod Partners With OpenCV to Empower the Next Gen of AI Builders," but the text available here contains no details about the partnership itself, such as its scope, terms or timing.
