NVIDIA / NVIDIA/TransformerEngine
2D by 2D per-block quantization in FP8 GEMM
Open
Nobody has claimed this yet.
enhancement
- Dominant language
- Python
- Stars
- 3.5k
- Forks
- 831
- Avg merge
- 3d 11h
- Merged PRs (30d)
- 65
Description
Is your feature request related to a problem? Please describe.
Float8BlockScaling does not support 2D by 2D quantization. I think this is super nice to have because 2D by 2D quant can reduce the number of quantization needed for output and Doutput.
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 transformer_engine/common/recipe/init.py at the Float8BlockScaling definition linked in the issue. Read the surrounding scaling configuration and related FP8 GEMM paths to determine the changes needed for 2D-by-2D quantization; done means Float8BlockScaling supports that mode and its behavior is covered by the relevant tests.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 25/100