Reproduce “Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mahmoud Ashraf
Flash RNN-T reduces training-memory use for streaming speech recognition by tiling valid alignment positions and recomputing hidden states and vocabulary logits during backpropagation. It retains blank and next-label probabilities, evaluating the full alignment space and vocabulary. On the H100 medium workload, it reaches 2,839 samples/s at 1.84 GiB; tiled runs also fit the long workload under retain scheduling where tested untiled alternatives run out of memory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.