NVIDIA / NVIDIA/TransformerEngine
Graph-Safe FP8 BlockScaling Support on Blackwell for GroupedLinear
Nobody has claimed this yet.
- 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.
We have Graph-Safe FP8 BlockScaling Support on Hopper which should technically work on Blackwell. We need the FP8 BlockScaling GroupedGEMM support on Blackwell to unblock this.
Describe the solution you'd like
General GEMM also has FP8 Blockscaling support on Blackwell. It does so why Broadcasting 128128 scales onto 321, 1*32 scaling factors needed for MXPF8 tensors. And does a MXFP8 GEMM.
We need to extend the same idea with GroupedGEMM
cc: @ptrendx @denera
Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.
Additional context
Add any other context or screenshots about the feature request here.
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
No files or tests are named. Start by tracing the existing Hopper Graph-Safe FP8 BlockScaling path and the Blackwell General GEMM implementation, then compare how GroupedGEMM handles MXFP8 scaling. Done means GroupedGEMM supports the requested Blackwell FP8 BlockScaling behavior without breaking existing paths.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 42/100