pytorch - 💡(How to fix) Fix [inductor] Support pre/post kernel execution hooks in compiled graphs [1 participants]

TL;DR

Implement a new mechanism to instrument kernel calls for profiling/tracing purposes, allowing user callbacks before and after each node execution with access to node name and input/output tensors.

Guidance

• Review the existing generate_intermediate_hooks function to understand its limitations and why it cannot be repurposed for profiling/tracing use cases.
• Consider implementing a new hook mechanism specifically designed for profiling/tracing, allowing for user callbacks around kernel calls.
• Evaluate the working implementation on the node-hooks branch, ensuring it meets the requirements for instrumenting kernel calls and providing access to node name and input/output tensors.
• Discuss with the maintainers and contributors to determine the best approach for integrating the new hook mechanism into the existing codebase.

Notes

The implementation details of the new hook mechanism are not specified, and the working implementation on the node-hooks branch should be reviewed to ensure it meets the requirements.

Recommendation

Apply workaround by implementing a new hook mechanism, as repurposing the existing generate_intermediate_hooks function may conflate different use cases and break existing users.