[Easy/Profiler] Set Duration to -1 for unfinished CPU events (#150131)

sraikund16 · pytorchmergebot · commit 8b043649147c · 2025-03-28T00:29:22.000Z
Summary: Some OSS Kineto users were requesting that we allow for 0 duration events in Kineto even though they won't be seen on the trace. To allow this we changed the handling of said events in D71510383. However this causes unfinished events in collection to never be post processed; this diff fixes said issue. Test Plan: https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree/traces/dynocli/0/1743102222/localhost/libkineto_activities_631490.json.gz&bucket=gpu_traces Differential Revision: D71993609 Pull Request resolved: #150131 Approved by: https://github.com/chuanhaozhuge, https://github.com/xw285cornell
diff --git a/torch/csrc/profiler/collection.cpp b/torch/csrc/profiler/collection.cpp
@@ -875,11 +875,11 @@ void passEventsToKineto(
   // Generate Kineto events for each event recorded by the PyTorch profiler.
   for (const auto i : c10::irange(results.size())) {
     const auto& e = results[i];
-    // (TODO): This is a temporary fix for async traces to make sure that we do
-    // not use int64 MIN as end time in Kineto. If we use that value, the
-    // duration will overflow and become a very large positive number. For a
-    // long term solution, add guards in kineto for each activity type
-    int64_t act_end_time = std::max(e->endTimeNS(), e->start_time_ns_);
+    // Here we are essentially setting the duration to -1 if the event never
+    // ends. This way Kineto will extend the event to the end of the trace. This
+    // is useful so that we can still have 0 duration events if necessary
+    // without extension
+    int64_t act_end_time = std::max(e->endTimeNS(), e->start_time_ns_ - 1);
     std::string name = e->name();
     if (!e->overload_name().empty()) {
       name = fmt::format("{}.{}", e->name(), e->overload_name());
