tangermeme adds PackedSequence: 2 bits per base, 16x memory savings and 9x faster shuffling
jmschreiber91 · x · 2026-10-20
tangermeme, a deep learning toolkit for genomics, merged PR #120 adding PackedSequence, which stores four-character alphabet sequences at 2 bits per position — 16x less memory than int8 one-hot encoding.
- Motivation: one-hot spends four values per position (4 bytes in int8, 16 in float32), which balloons for genome-wide windows, dinucleotide shuffles, and deepliftshap references reused across models
- Compatibility: one-hot calls are unchanged, bit-identical outputs, no timing difference; packing activates only when packed input or returnpacked=True is used
- Results: dinucleotide shuffling of 10k windows in packed space drops from 7.86s to 0.85s and 8.7GB to 0.7GB
The branch also underwent two independent adversarial reviews.
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