STOF framework optimizes sparse Transformer on GPU via analytical kernel mapping for MHA and two-stage search for fusion, reporting up to 1.6x MHA and 1.4x end-to-end speedups over prior work.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Accelerating Sparse Transformer Inference on GPU
STOF framework optimizes sparse Transformer on GPU via analytical kernel mapping for MHA and two-stage search for fusion, reporting up to 1.6x MHA and 1.4x end-to-end speedups over prior work.