NVIDIA / NVIDIA/TransformerEngine
did not get improvement from tp/sp overlap
Open
Nobody has claimed this yet.
bug
- Dominant language
- Python
- Stars
- 3.5k
- Forks
- 831
- Avg merge
- 3d 11h
- Merged PRs (30d)
- 65
Description
i run examples of te_comm_gemm_overlap.py, and remove backwards code, only forward code.
compared to tp allreduce, the tp/sp allgather + reduce scatter is slower
my gpu is 8 h20
is there any other args needs to change?
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with te_comm_gemm_overlap.py and run the forward-only comparison between TP allreduce and TP/SP allgather plus reduce-scatter on the reported 8 H20 GPUs. Inspect the example's available arguments and determine whether a configuration change explains the slower result; done means identifying the relevant setting or documenting why the overlap does not improve performance.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100