method · first seen 16 Oct 2026
memory-efficient loss computation
Techniques such as fused kernels and tiled recomputation that reduce peak memory for large intermediate tensors in training.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss ComputationPapers · 16 Oct 2026 · Mahmoud Ashraf
Latest sources
- Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss ComputationPapers · 16 Oct 2026 · Mahmoud Ashraf