Announcement
ORL claims similar performance at much higher speed than speculative decoding
The ORL team says co-designing algorithms, its training and inference engine, and kernels improves the speed–intelligence trade-off.
TLDR
The ORL team argues that diffusion language models can improve the trade-off between speed and intelligence. It claims ORL achieves similar performance much faster than speculative decoding through co-design of its algorithms, training and inference engine, and kernels. The team also says ORL-DLM is better than “lossy spec” and links to live demos.
Combined views
1K
2 Sources, first seen ago
