[CUDA] hack for host-memory stats test #163119
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/163119
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 67564ab with merge base d4554bc ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
which of the stats is failing? It's doing empty_cache so most should be fine, and we can relax checks on those that are failing rather than resort to ugly workarounds |
|
The failure signature I'm seeing most often is expecting 0 bytes host (pinned memory) and seeing 76 bytes allocated |
|
Fine, that means some other tests allocate and hold on to some memory, we can rewrite this test to be more robust to initial allocated memory (check that e.g. allocation is small in the beginning, and that it returns to this value in the end) |
Basically this test appears to fail when run with other tests that manipulate pinned memory, even with the
@serialTestdecoratorAn ugly workaround here is to ensure it runs first due to alphabetical ordering...
cc @ptrblck @msaroufim @jerryzh168