intel / intel/intel-extension-for-tensorflow
Memory leak.
Open
aitce
bug
- Dominant language
- C++
- Stars
- 353
- Forks
- 45
- PR merge metrics
- No merged PRs in 30d
Description
I have been training using tensorflow + keras on nvidia GPU for a while.
Recently I experimented with A770. With some efforts, I finally got it working, except that there is a memory leak.
The same code works fine on nvidia 3090, it uses about 8GB memory, very stably.
With A770, it starts with 8GB and grows very quickly until killed because of OOM.
I used tracemalloc to see where is the leak. No luck. So it's not in python code.
I haven't got time to get more details of it...
Contributor guide
Assessment
This issue has not been assessed yet.