Blackwell / GB10 support for MiniMax MSA decode-sparse path via HF Kernel Hub
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 422
- Forks
- 57
- Avg merge
- 15h 50m
- Merged PRs (30d)
- 1
Description
Hi MiniMax team,
First, thank you for the great work on MSA and for open-sourcing it. Also thanks for contributing the original HF Kernel Hub package.
I tested and packaged a FlashRT-maintained Blackwell / GB10 extension of the MiniMax MSA decode-sparse path for the HF Kernel Hub:
https://huggingface.co/kernels/flashrt/MiniMaxAI-msa-blackwell
This package targets NVIDIA Blackwell compute capability 12.x and has been validated on DGX Spark / GB10 / SM121. The main goal is to make the MiniMax MSA decode-sparse path usable beyond the original SM100-only target.
Source repo:
https://github.com/LiangSu8899/FlashRT-HF-kernels/tree/main/MiniMaxAI-msa-blackwell
I’d be happy to keep iterating on this and gradually port more of the SM100 MSA API surface to Blackwell where it makes sense.
Contributor guide
No contributing guide indexed for this repository
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
Review the linked Hugging Face Kernel Hub package and the MiniMaxAI-msa-blackwell source repository first. Clarify which Blackwell support should be integrated into MSA, which SM100 API surface is in scope, and what validation on Blackwell or GB10 would define completion.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100