[Feature]: Support for FlashMLA #13735
Description
🚀 The feature, motivation and pitch
https://github.com/deepseek-ai/FlashMLA
Achieving up to 3000 GB/s in memory-bound configuration and 580 TFLOPS in computation-bound configuration on H800 SXM5, using CUDA 12.6.
Alternatives
No response
Additional context
No response
Before submitting a new issue...
- Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.