FlashQLA v0.1.2
🚀 Highlights
-
Flash-Linear-Attention backend integration by @Erix025 — FlashQLA can now be used as a drop-in backend for FLA's GDN (chunked gated delta rule) via the flash-linear-attention interface.
-
SM120 forward pass contributed by @minatoyukinaa — added Blackwell SM120 forward kernels with chunk_size=32 and efficient shared memory reuse.
🛠️ Additional Improvements
- Fixed a critical issue caused by id-base tensor cache. Reported by @celiolarcher.
📊 Benchmarks
🙏 Acknowledgements
Special thanks to @Erix025, @minatoyukinaa, and @celiolarcher for their contributions, and to all community members for their valuable feedback. We welcome your continued participation through issues and pull requests.
Contributors
celiolarcher, Erix025, and minatoyukinaa