-
Notifications
You must be signed in to change notification settings - Fork 2.7k
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
Buffer donation to a jit function on GPU #1273
Comments
Buffer donation has been checked in! |
Thanks Peter, do you know how can I leverage it to reduce the memory consumption in the example above? So far, even if I do
I still get peak memory of 4x2 = 8Gb, and a message
|
I believe that means that there wasn't an output with the same shape that could have reused that buffer (or there weren't an equal number of such outputs as inputs). |
Interesting - how come it doesn't work in this example then? From my understanding here there's 1 input, 1 output, both of shape and type |
FYI buffeer donation is only supported on TPU at the moment, XLA team are working to support this on CPU/GPU but that may be why we cannot use the donation. |
I see, thanks! Could you please reopen this issue then? |
Fixed by #3800 |
Below is a CNN iteratedly applied to a 2Gb input. It produces a 4x2Gb = 8 Gb peak memory consumption.
Without JIT, the peak memory consumption is 2x2Gb = 4 Gb, as is expected.
Would be great to achieve a comparable memory usage with JIT by input buffer donation to the jit function (not sure on the exact terminology).
Thanks a lot!
The text was updated successfully, but these errors were encountered: