#Device pass through

1 messages · Page 1 of 1 (latest)

opal dove
#

Haven't tried it yet, but you mean that with dagger I should be able to pass through the FPGA device?

naive zealot
#

Dagger doesn't support that today, but I'm saying it would be cool if it did 🙂

#

There's a related design discussion, which is to pass through native GPU from the client to the container, using webgpu. That would mean for example, allowing a container to use the Apple Silicon from a Mac client.

As you can imagine there is a lot of demand for that, with generative AI workloads being all the rage. I'm thinking once we ship that (no timeline at the moment), other devices could benefit.

steady plank
#

This is the same feature needed in modules to support KVM for cross os modules (eg windows/macos in Dagger), as well as GPU from non Nvidia vendors, and accelerator network cards used in high performance computing. I’m hyped to hear you think adding this would be cool.

CC: @shell egret

naive zealot
steady plank
#

Here the guy is talking about FPGAs

#

Though webgpu is another cool use case

naive zealot
#

yes I know, I'm pointing out a potential difference with past discussions about device access, which is that the device would not be mounted into the container from the buildkit/runc host

#

I love the clean sandboxing of client streaming. Not sure if it works for all devices. But it would be a fantastic way to separate concerns between "dumb" container runners and "smart" clients

steady plank
#

Has anyone actually tested webgpu as a backend for megatron deepspeed or the other AI frameworks. My impression talking to people who run those workloads was you need actual Cuda, HIP, or OneAPI for that to work

naive zealot
steady plank