A fused tensor-core sparse attention kernel (SDDMM, softmax, SpMM) that achieves large speedups over prior baselines on H100 and A30.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Fused3S: Fast Sparse Attention on Tensor Cores
A fused tensor-core sparse attention kernel (SDDMM, softmax, SpMM) that achieves large speedups over prior baselines on H100 and A30.