pytorch - 💡(How to fix) Fix Make torchrun options ( --nnodes, --nproc-per-node, etc.) same a mpirun/mpiexec [1 comments, 2 participants]

TL;DR

Modify the common options to match the syntax and functionality of mpirun/mpiexec, specifically incorporating options like -n and -ppn.

Guidance

• Review the documentation and implementation of mpirun/mpiexec to understand the expected behavior of options like -n and -ppn.
• Identify the current implementation of common options in the codebase and plan modifications to align with the mpirun/mpiexec syntax.
• Consider the impact of these changes on existing users and workflows, potentially requiring updates to documentation or user interfaces.
• Test the modified options thoroughly to ensure compatibility and correctness.

Notes

The lack of specific technical details in the issue limits the ability to provide a precise fix or code example. The guidance provided is based on the implied requirement to match the functionality of mpirun/mpiexec.

Recommendation

Apply workaround: Modify the common options to match mpirun/mpiexec syntax, as this directly addresses the stated goal without implying a version upgrade or other changes.