Speeding up DeepLIFT/SHAP: packed-sequence dinucleotide shuffling avoids memory blowup
jmschreiber91 · x · 2026-10-20
While speeding up a DeepLIFT/SHAP implementation, the author hit a bottleneck: dinucleotide shuffling each input independently was slow, but shuffling all sequences together (expecting 100k+ sequences, 20x shuffles each) caused an insane memory blowup.
The fix is computing directly on packed sequences — tangermeme functions now support ersatz operations directly on packed data, so transforms like shuffling no longer require expanding into huge tensors.
More from Research
- Gojo project generates 108 ornament designs from 36 skeletons with hierarchical pattern-aware coloring — debreuil · 2026-10-20
- Paper: 'Before They Can Solve' predicts a base model's post-training potential early — hyunw_kim · 2026-10-20
- Base-model probes predict cross-benchmark coding-agent potential at ρ=0.976 — hyunw_kim · 2026-10-20
- Three probes track post-training SWE-bench rankings, BPB hits ρ=0.964 — hyunw_kim · 2026-10-20
- NVIDIA asks: must you post-train every base model to find the best coding agent? — hyunw_kim · 2026-10-20
- SaTML got 450 submissions, 2x last year — chairs early-rejected ~40% of them — m2saxon · 2026-10-20