[Feature] Support python-based mscclpp-test
Binyang2014 opened this issue · 0 comments
Binyang2014 commented
We've developed mscclpp-test using C++ and CUDA. However, adjusting parameters for various scenarios has proven challenging. Introducing a Python-based version of mscclpp-test could simplify this process.
Parameters need to tune:
- number of thread blocks
- number of thread per block
- pipeline number for multi-nodes collective