NVIDIA / NVIDIA/TransformerEngine
[BUG] fp8blockscaled training not converging on sm120
Open
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 3.5k
- Forks
- 831
- Avg merge
- 3d 11h
- Merged PRs (30d)
- 65
Description
I have tried nvfp4 training which converges on sm120, but fp8blockscaled recipe won't converge for any of its available options. Is it because of power of 2 scale (cannot be set False for sm120 arch) that causes the problem?
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
The issue names no file, test, or entry point. Begin by reproducing fp8blockscaled training on sm120 across its available options, then determine whether the power-of-two scale setting explains non-convergence; done means the cause is established and a validated fix or documented limitation is available.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100