-
Notifications
You must be signed in to change notification settings - Fork 6.8k
Conversation
Hey @leezu , Thanks for submitting the PR
CI supported jobs: [miscellaneous, website, centos-gpu, centos-cpu, sanity, unix-cpu, unix-gpu, windows-gpu, edge, windows-cpu, clang] Note: |
@mxnet-bot run ci [centos-gpu, windows-gpu] |
Jenkins CI successfully triggered : [centos-gpu, windows-gpu] |
@Jerryzcn I think this is also related to your previous benchmark. |
@ciyongch should we backport this to 1.7? |
Hi @leezu , if the issue also appears in 1.7 then please help to backport to 1.7 and 1.x branches and tag me to the new PR, thanks! |
This issue is present in all versions of Gluon. OK, let's backport the fix. |
Fix leak of ndarray objects in the frontend due to reference cycle. Backport of 3e676fc
Fix leak of ndarray objects in the frontend due to reference cycle. Backport of 3e676fc
@ciyongch I created the backport PRs |
Thanks @leezu to help backport the PR. |
@ChaiBapchya @access2rohit This may have also fixed our out-of-memory issue in large tensor nightly test when running them in sequence. |
Fix leak of ndarray objects in the frontend due to reference cycle.
This reverts commit c4d9270.
This reverts commit b523527.
Fix leak of ndarray objects in the frontend due to reference cycle.
Fix leak of ndarray objects in the frontend due to reference cycle. Backport of 3e676fc
After this commit apache/mxnet#18328 , some memory leak were fixed. Whitout this commit faster rcnn traning cannot be successfully closed. These commits can be commit again after this yolo training fix. apache/mxnet#18692 apache/mxnet@0496690
* Fix yolo to support a memory leak fix After this commit apache/mxnet#18328 , some memory leak were fixed. Whitout this commit faster rcnn traning cannot be successfully closed. These commits can be commit again after this yolo training fix. apache/mxnet#18692 apache/mxnet@0496690 * fix all generator error in windows when training with multiprocessing * add pylint disable not-callable * Fix pylint * Fix pylint Co-authored-by: Joshua Z. Zhang <cheungchih@gmail.com>
Description
Previously the _BlockScope keeps references to the parameter ndarrays, preventing memory from being freed if a Block is not used anymore. Among other problems, this causes memory usage to increase constantly in unittests (due to testing different blocks and disposing them at the end of the test) until the garbage collector kicks in (which can be too late and the system can run OOM as the parameter arrays can be large).