[REQUEST] Support running GPU kernels with arrays in system memory on Grace Hopper
Open
@shi-eric is already working on this.
Since Apr 29, 2026.
feature request
platform
- Dominant language
- Python
- Stars
- 7.1k
- Forks
- 624
- Avg merge
- 3d 17h
- Merged PRs (30d)
- 5
Description
Description
It should be possible to mix-and-match arrays on Grace Hopper systems, e.g. CPU kernels can access arrays residing in GPU memory and vice versa. A runtime check requiring all arrays to be allocated on the same device as the kernel is one of the things preventing this from working:
There might be additional changes required for this capability to be performant.
Similarly, it should be possible to mix-and-match GPU arrays when peer access is enabled on peer-capable GPUs.
Context
Take advantage of platform features.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.