Open jerryyangli opened 5 years ago
The main dependency, OpenAI's blocksparse library, has no PyTorch bindings. https://github.com/openai/blocksparse/issues/2 So that would be a starting point, unless an alternative to blocksparse has come out in the last 1.5 years.
Is it not possible to use this approach ? Seems pretty feasible if I'm not missing anything. Thoughts @jonasschneider @scott-gray ?
Is it possible to release a PyTorch implementation of the method?