sraikund16
sraikund16
Summary: In the recent weeks, we have encountered bugs in both the normal synchronous trace and on-demand tracing. This diff on its own does sanity checking to make sure the...
Summary: github issue: https://github.com/pytorch/pytorch/issues/73828 Whenever we transition from record and save to warmup, we instantiate a new backend profiler which wipes out the last cycles information. We should keep the...
Summary: Users have complained that STAGE level logs are print by default. If a user is running many profiles it can certainly clutter STDOUT. Lower the STAGE level such that...
Summary: As reported by this issue: https://github.com/pytorch/pytorch/issues/83584 We already store the entries in evt.stack so there is no need to cap the limit when we output the table to 5...
Summary: Recently, we have had users seen empty traces when the system is idle leading to confusion as to whether it was caused by a bug in kineto formatting or...
After upgrading the CUDA image to 12.4 we are having segfault failures in the following tests: 19 - CuptiRangeProfilerApiTest.asyncLaunchUserRange (SEGFAULT) 20 - CuptiRangeProfilerApiTest.asyncLaunchAutoRange (SEGFAULT) 24 - CuptiRangeProfilerTest.UserRangeTest (SEGFAULT) 25 -...
Summary: Right now we align Roctracer events to system clock blindly regardless of what we are using in torch.profiler. We should use a clock based on what is defined instead....
Summary: Add debug to show that some events in roctracer can start before the previous one ends Differential Revision: D63033163
Differential Revision: D61943373
### Problem Description We notice that many of the events in Roctracer for a single GPU and single queue have a "tie". The first event ends at the exact same...