-
-
Notifications
You must be signed in to change notification settings - Fork 606
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
handle data movement with MLDataDevices.jl #2492
Conversation
I assume the error is due to Enztme not yet supporting 1.11, does it work on the LTS? |
Ok I think this is ready to go. Buildkite passes on CUDA but fails on Metal and AMDGPU, due to the fact that it seems to be failing to load the preference. |
ah ok, I didn't know. It works on 1.10 indeed. |
Maybe also adding an LTS cI run would be helpful here? (Now that it’s released, were starting to adapt to the new intrinsics and array operators [this was sadly for us a more significant set of changes]) |
I kind of imagine that happened for most other packages too since the release just happened the other day |
Should probably replace |
One note for the follow-up: the MLDataDevices path does not flip conv kernels as is required for the AMDGPU backend: Flux.jl/ext/FluxAMDGPUExt/functor.jl Lines 87 to 109 in aa035e9
We should add this functionality before cutting any releases. |
After FluxML/Flux.jl#2492 also Flux relies on MLDataDevices.
After FluxML/Flux.jl#2492 also Flux relies on MLDataDevices.
Implements the move to MLDataDevices.jl. Closes #2482
Thanks to the extended functionality given by MLDataDevices.jl we can also easily fix #2490
The
gpu/cpu
logic could also be offloaded to MLDataDevices.jl, will do that in a followup PR.Unfortunately this PR has become rather large also due to some unrelated changes to fix the CI since it was broken due to:
Worth openining separate Issues to keep track of these.
PR Checklist