vllm - 💡(How to fix) Fix [vLLM IR] Propagate IR op names to torch profiler annotations [1 participants]

TL;DR

Annotating custom and codegen-ed Triton kernels with corresponding IR op and implementation name can improve the readability of torch profiles.

Guidance

• Identify the source of kernel generation to determine where annotations can be added, potentially in the codegen process or when creating custom kernels.
• Investigate existing profiling tools like nsys for any built-in features or APIs that support custom annotations.
• Review related issues like #39215 for potential insights or partial solutions that could be adapted.
• Consider modifying the kernel creation process to include metadata about the corresponding IR operation and implementation name.

Example

No specific code example can be provided without more context on the kernel generation process.

Notes

The feasibility of annotating kernels may depend on the specific implementation details of the torch profiling tool and the Triton kernel generation process, which are not provided in the issue.

Recommendation

Apply workaround: Modify the kernel creation process to include annotations, as this seems to be a targeted solution to the problem of identifying kernel operations in profiles.