Event Handler

The Event Handler module in AccelProf is responsible for capturing and translating runtime events from multiple sources:

• High-level Deep Learning (DL) frameworks (e.g., PyTorch)

  • Example callback: c10::reportMemoryUsage

• Low-level instrumentation APIs

  • Example: SANITIZER_CBID_LAUNCH_BEGIN

These events are then processed by a modular set of handler functions such as:

• AccelProf::tensor_call_handler for tensor allocations

• AccelProf::kernel_call_handler for kernel launch events

This abstraction enables AccelProf to support diverse sources of profiling data while maintaining a unified internal event format.